Image processing model training, image processing method

Abstract

The embodiments of the present specification provide image processing model training and image processing methods. The image processing model training method comprises the following steps: determining target training data and augmented training data; using an initial image processing model to respectively perform object detection on a sample image and a flipped sample image, to obtain a detection result of the sample image and a flipped detection result of the flipped sample image; determining predicted object region information of a target object in the sample image based on the detection result and the flipped detection result; and performing model training on the initial image processing model based on the object region information, the predicted object region information, the detection result, a sample label, the flipped detection result and a flipped sample label, to obtain a trained target image processing model, so that a target image processing model capable of more accurate object detection is obtained.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of artificial intelligence technology, and in particular to an image processing model training method; one or more embodiments of this specification also relate to another image processing model training method, an image processing method, another image processing method, an information processing method, a computing device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] With the continuous development of computer technology and artificial intelligence technology, data processing operations can be performed through neural network models, thereby improving the efficiency of data processing; for example, object detection operations can be performed on images through neural network models.

[0003] Current neural network models suffer from inaccurate detection results when performing object detection on images. Therefore, obtaining accurate detection results through neural network models has become an urgent problem to be solved. Summary of the Invention

[0004] In view of the above, embodiments of this specification provide an image processing model training method. One or more embodiments of this specification also relate to another image processing model training method, an image processing method, another image processing method, an information processing method, a computer-aided diagnosis method for cancer, a computer-aided diagnosis method for nasopharyngeal carcinoma, a computer-aided diagnosis method for breast cancer, an image processing model training device, another image processing model training device, an image processing device, another image processing device, an information processing device, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, an image processing model training method is provided, comprising:

[0006] Determine the target training data and the augmented training data, wherein the target training data includes a sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object; and the augmented training data includes a flipped sample image corresponding to the sample image and a flipped sample label of the flipped sample image.

[0007] Using the initial image processing model, object detection is performed on the sample image and the flipped sample image respectively to obtain the detection results of the sample image and the flipped detection results of the flipped sample image;

[0008] Based on the detection results and the flip detection results, the predicted object region information of the target object in the sample image is determined;

[0009] Based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label, the initial image processing model is trained to obtain the trained target image processing model.

[0010] According to a second aspect of the embodiments of this specification, an image processing model training apparatus is provided, comprising:

[0011] The training data determination module is configured to determine target training data and augmented training data. The target training data includes a sample image of a target object located at a target location, sample labels of the sample images, and object region information of the target object. The augmented training data includes a flipped sample image corresponding to the sample image and a flipped sample label of the flipped sample image.

[0012] The object detection module is configured to use an initial image processing model to perform object detection on the sample image and the flipped sample image respectively, and obtain the detection result of the sample image and the flipping detection result of the flipped sample image;

[0013] The region information determination module is configured to determine the predicted object region information of the target object in the sample image based on the detection result and the flip detection result;

[0014] The model training module is configured to train the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label, so as to obtain the trained target image processing model.

[0015] According to a third aspect of the embodiments of this specification, another image processing model training method is provided, comprising:

[0016] The target training data and the augmentation training data are determined. The target training data includes a sample CT image of a target tumor located in the nasopharynx, a sample label of the sample CT image, and tumor region information of the target tumor. The augmentation training data includes a flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image.

[0017] Using an initial image processing model, tumor detection is performed on the sample CT image and the flipped sample CT image respectively to obtain the tumor detection results of the sample CT image and the flipped tumor detection results of the flipped sample CT image.

[0018] Based on the tumor detection results and the flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined;

[0019] Based on the tumor region information, the predicted tumor region information, the tumor detection result, the sample label, the flipped tumor detection result, and the flipped sample label, the initial image processing model is trained to obtain the trained target image processing model.

[0020] According to a fourth aspect of the embodiments of this specification, another image processing model training apparatus is provided, comprising:

[0021] The training data determination module is configured to determine target training data and augmentation training data. The target training data includes a sample CT image of a target tumor located in the nasopharynx, a sample label of the sample CT image, and tumor region information of the target tumor. The augmentation training data includes a flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image.

[0022] The tumor detection module is configured to use an initial image processing model to perform tumor detection on the sample CT image and the flipped sample CT image respectively, and obtain the tumor detection result of the sample CT image and the flipped tumor detection result of the flipped sample CT image.

[0023] The region information determination module is configured to determine the predicted tumor region information of the target tumor in the sample CT image based on the tumor detection results and the flipped tumor detection results.

[0024] The model training module is configured to train the initial image processing model based on the tumor region information, the predicted tumor region information, the tumor detection result, the sample label, the flipped tumor detection result, and the flipped sample label, to obtain the trained target image processing model.

[0025] According to a fifth aspect of the embodiments of this specification, an image processing method is provided, comprising:

[0026] Determine the image to be detected of the target object located at the target location;

[0027] Object detection is performed on the image to be detected using a target image processing model to obtain the detection result of the image to be detected. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by performing object detection on the sample image using the initial image processing model, and the flip detection result is determined by performing object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0028] According to a sixth aspect of the embodiments of this specification, an image processing apparatus is provided, comprising:

[0029] The image determination module is configured to determine the image to be detected of a target object located at the target location.

[0030] An object detection module is configured to perform object detection on the image to be detected using a target image processing model to obtain the detection result of the image to be detected. The target image processing model is obtained by training an initial image processing model based on object region information determined from target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by performing object detection on the sample image using the initial image processing model, and the flip detection result is determined by performing object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0031] According to a seventh aspect of the embodiments of this specification, another image processing method is provided, applied to a client of a medical system, comprising:

[0032] In response to the user's selection operation on the client's display interface, the medical image to be detected is determined;

[0033] The medical image to be detected is sent to the server of the medical system, and the detection result returned by the server is received. The detection result is the output result after object detection of the medical image to be detected by the target image processing model. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection result, sample label, flip detection result, and flip sample label. The target training data includes the sample image of the target object located at the target site, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by object detection of the sample image using the initial image processing model. The flip detection result is determined by object detection of the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0034] The detection results are displayed to the user through the display interface.

[0035] According to an eighth aspect of the embodiments of this specification, another image processing apparatus is provided, applied to a client of a medical system, comprising:

[0036] The image determination module is configured to determine the medical image to be detected in response to a user's selection operation on the client's display interface;

[0037] An object detection module is configured to send the medical image to be detected to the server of the medical system and receive the detection result returned by the server. The detection result is the output of object detection on the medical image to be detected using a target image processing model. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by object detection on the sample image using the initial image processing model, and the flip detection result is determined by object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0038] The results display module is configured to display the detection results to the user through the display interface.

[0039] According to a ninth aspect of the embodiments of this specification, an information processing method is provided, comprising:

[0040] Identify the CT image of the target tumor located in the nasopharynx;

[0041] A target image processing model is used to perform tumor detection on the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by performing tumor detection on the sample CT image using the initial image processing model. The flipped tumor detection result is determined by performing tumor detection on the flipped sample CT image using the initial image processing model. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0042] The information processing result is determined based on the tumor detection result of the CT image to be detected.

[0043] According to a tenth aspect of the embodiments of this specification, an information processing apparatus is provided, comprising:

[0044] The image determination module is configured to determine the CT image to be detected for a target tumor located in the nasopharynx.

[0045] The object detection module is configured to perform tumor detection on the CT image to be detected using a target image processing model, and obtain the tumor detection result of the CT image to be detected.

[0046] The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0047] The information processing module is configured to determine the information processing result based on the tumor detection result of the CT image to be detected.

[0048] According to the eleventh aspect of the embodiments of this specification, a computer-aided diagnosis method for cancer is provided, comprising:

[0049] Identify the CT image of the target tumor located at the target site;

[0050] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located at the target site, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0051] According to the twelfth aspect of the embodiments of this specification, a computer-aided diagnostic method for nasopharyngeal carcinoma is provided, comprising:

[0052] Identify the CT image of the target tumor located in the nasopharynx:

[0053] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0054] According to a thirteenth aspect of the embodiments of this specification, a computer-aided diagnostic method for breast cancer is provided, comprising:

[0055] Identify the CT image of the target tumor located in the breast region;

[0056] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the breast region, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0057] According to a fourteenth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0058] Memory and processor;

[0059] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of any of the above methods.

[0060] According to a fifteenth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0061] According to a sixteenth aspect of an embodiment of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0062] This specification provides an image processing model training method through one or more embodiments. This method utilizes a large amount of target training data and augmented training data to train an initial image processing model, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on these detection and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Subsequently, based on this target image processing model, accurate detection results are obtained during object detection of images. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the application of an image processing method provided in one embodiment of this specification;

[0064] Figure 2 This is a flowchart illustrating an image processing model training method provided in one embodiment of this specification;

[0065] Figure 3 This is a schematic diagram of a CT scan in an image processing model training method provided in one embodiment of this specification;

[0066] Figure 4 This is a flowchart illustrating the processing procedure of an image processing model training method provided in one embodiment of this specification.

[0067] Figure 5 This is a flowchart of another image processing model training method provided in one embodiment of this specification;

[0068] Figure 6 This is a flowchart illustrating an image processing method provided in one embodiment of this specification;

[0069] Figure 7 This is a flowchart of another image processing method provided in one embodiment of this specification;

[0070] Figure 8 This is a flowchart illustrating an information processing method provided in one embodiment of this specification;

[0071] Figure 9 This is a schematic diagram of the structure of an image processing model training device provided in one embodiment of this specification;

[0072] Figure 10 This is a schematic diagram of the structure of another image processing model training device provided in one embodiment of this specification;

[0073] Figure 11 This is a schematic diagram of the structure of an image processing apparatus provided in one embodiment of this specification;

[0074] Figure 12 This is a schematic diagram of the structure of another image processing apparatus provided in one embodiment of this specification;

[0075] Figure 13 This is a schematic diagram of the structure of an information processing device provided in one embodiment of this specification;

[0076] Figure 14 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0077] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0078] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0079] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0080] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0081] In one or more embodiments of this specification, a large model refers to a deep learning model with a large number of model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. A large model can also be called a foundation model. It is pre-trained using large-scale unlabeled corpora to produce a pre-trained model with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0082] In practical applications, large models only require a small number of samples to fine-tune the pre-trained model before they can be applied to different tasks. Large models can be widely used in fields such as Natural Language Processing (NLP) and Computer Vision. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. The main application scenarios for large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0083] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0084] Nasopharyngeal carcinoma (NPC) is an epithelial malignant tumor originating from the nasopharyngeal mucosa. NPC is one of the most common cancers of the nasopharynx, and timely diagnosis and treatment can significantly improve its prognosis. Compared with other head and neck cancers, NPC has unique clinical characteristics and is extremely sensitive to radiotherapy (RT). Therefore, most NPC patients can be cured with radiotherapy; one of the most important tasks in radiotherapy planning is to delineate the gross tumor volume (GTV) in the planning CT (pCT) scan (without iodine contrast agent injection).

[0085] pCT (planning computed tomography) is an advanced imaging technology widely used in the medical field.

[0086] GTV (Gross Tumor Volume): refers to the gross tumor volume.

[0087] RT (Radiation therapy): refers to radiotherapy.

[0088] MRI (Magnetic Resonance Imaging): refers to magnetic resonance imaging.

[0089] Dice score (Dice coefficient): a method for evaluating segmentation performance.

[0090] Projection head: A projection head is a network structure used to extract features or representations. It is usually one or more fully connected layers, or it can be a convolutional layer, that maps the high-dimensional features of the input to a low-dimensional vector space to facilitate subsequent tasks.

[0091] DSC (Dice Similarity Coefficient): A metric used to evaluate the similarity of binary image segmentation results. The DSC loss function is often used in fields such as medical image segmentation because it is sensitive to the size of the overlapping parts and can effectively measure the degree of overlap between the predicted results and the true labels.

[0092] CE Loss (Cross-Entropy Loss): refers to the cross-entropy loss function.

[0093] Affine registration: an image processing technique commonly used to globally align medical images from different time points or different modalities.

[0094] Semantic symmetry is crucial for various medical image analyses, as human anatomical structures typically exhibit bilateral symmetry, such as the brain, breasts, lungs, and pelvis.

[0095] Margin Loss: Margin Loss is a loss function that defines loss by setting a "margin" or "margin". Its purpose is to create a certain gap between the predicted value of the correct class and the predicted value of the incorrect class.

[0096] SAT (Semantic Asymmetry Tumor): refers to a tumor with semantic asymmetry.

[0097] SATs (Semantic Asymmetry Tumor Segmentation) refer to semantically asymmetric tumor segmentation.

[0098] Siamese refers to Siamese Net, which is two identical CNNs (Convolutional Neural Networks) that share weights.

[0099] LeakyReLU is an improved ReLU function, which is a non-linear activation function with parameters.

[0100] ReLU (Rectified Linear Unit): The activation function is a commonly used mathematical function in artificial neural networks. It is used to introduce nonlinearity and help the network learn complex patterns and relationships in the data.

[0101] InstanceNorm: Short for Instance Normalization, it is a technique used in deep learning to accelerate training and improve model performance.

[0102] Epoch: refers to the process by which the model completely traverses and learns the entire dataset once.

[0103] With the continuous development of computer technology and artificial intelligence technology, data processing operations can be performed through neural network models, thereby improving the efficiency of data processing; for example, object detection operations can be performed on images through neural network models; however, current neural network models have the problem of inaccurate detection results when performing object detection on images.

[0104] For example, in radiotherapy for nasopharyngeal carcinoma (NPC), clinicians typically delineate the gross tumor volume (GTV) of the primary tumor on a non-contrast-guided computed tomography (pCT) scan to ensure accurate radiation dose delivery. This is a challenging task even for experienced radiation oncologists because NPC tumors can invade different types of surrounding tissues, creating poor contrast with the NPC tumor on the pCT (especially soft tissues such as mucous membranes, muscles, and nerves).

[0105] In clinical practice, diagnostic MRI is often used to guide manual GTV delineation in pCT. However, cross-modal registration between MRI and pCT is not straightforward due to significant modal gaps and variations in scan extent. Radiation oncologists typically use rigid transformations to roughly align MRI and pCT by matching skeletal structures such as the spine, or they simply open two viewers on a computer to simultaneously view the pCT and MRI and mentally fuse them to facilitate NPC GTV delineation. This process is very time-consuming (at least 1-2 hours per case) and error-prone. Therefore, accurate automatic segmentation of NPC GTV in pCT is essential; that is, because of the poor contrast between NPC tumors and many adjacent normal tissues, radiation oncologists need to manually delineate the GTV using diagnostic MRI, mentally aligning and fusing it. However, this process is both time-consuming and error-prone.

[0106] To address the aforementioned issues, this specification provides a method for automated NPC GTV segmentation based on deep learning. However, most MRI scans do not provide direct measurements of the electron density required for RT planning. Therefore, the MRI-derived tumor mask still needs to be spatially converted to pCT via image registration, which can lead to potential alignment errors (as large errors typically occur in cross-modal registration). Some studies have explored segmenting NPC GTVs only on enhanced CT scans (UNETR, SwinUNETR, and its variant SwinUNETR-v2). However, these typically yield relatively low performance, e.g., Dice score (DSC) below 70%. Only a few methods have solved the problem of automated NPC GTV segmentation using both CT and MRI. Observations have shown that multimodal segmentation performance may be inferior to single-modal segmentation performance if there is a large registration error between the paired CT and MRI scans.

[0107] Based on this, this specification provides an image processing model training method. One or more embodiments of this specification also relate to another image processing model training method, an image processing method, another image processing method, an information processing method, a computer-aided diagnosis method for cancer, a computer-aided diagnosis method for nasopharyngeal carcinoma, a computer-aided diagnosis method for breast cancer, an image processing model training device, another image processing model training device, an image processing device, another image processing device, an information processing device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0108] See Figure 1 , Figure 1 This diagram illustrates an application of an image processing method according to an embodiment of this specification. The client 102 identifies a target tumor located in the nasopharynx using a CT image and sends the CT image to a server 104. The server 104 uses an image segmentation model trained based on target training data and augmentation training data to perform tumor segmentation on the CT image, thereby obtaining a tumor segmentation map. This tumor segmentation map is then sent back to the client 102. This method significantly improves GTV segmentation in nasopharyngeal carcinoma by utilizing semantic symmetry in anatomical structures. The tumor segmentation map can potentially be used in radiotherapy practice to standardize GTV depiction of nasopharyngeal carcinoma and reduce the workload and variability for radiation oncologists.

[0109] See Figure 2 , Figure 2 A flowchart of an image processing model training method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0110] Step 202: Determine the target training data and the augmented training data, wherein the target training data includes a sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object, and the augmented training data includes a flipped sample image corresponding to the sample image and a flipped sample label of the flipped sample image.

[0111] The target training data and the augmented training data are training data used to train the initial image processing model, and the augmented training data are training samples obtained by performing sample augmentation processing on the target training data.

[0112] The target site can be a specific part of the target organism, such as the nasopharynx, chest, lungs, mammary glands, pelvis, or brain; the target organism can be a human or an animal; it should be noted that the target site can be a part that satisfies semantic symmetry. The target object can be understood as a target tumor, diseased tissue, diseased area, or cancerous area. See details in [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of a CT scan in an image processing model training method provided in one embodiment of this specification. Figure 3 Three CT scan images of the nasopharynx are shown. The area enclosed by the white dashed line is the tumor area. Each CT scan image has a central sagittal axis, and the nasopharynx is symmetrical about the left and right sides along the central sagittal axis.

[0113] It should be noted that in one or more embodiments of this specification, the application of the image processing model training method in the nasopharyngeal carcinoma scenario is used as an example to explain the image processing model training method. The image processing model training method provided in this specification can be applied to cancer scenarios, breast cancer scenarios, or pelvic cancer scenarios. For the specific implementation methods for cancer, breast cancer, and pelvic cancer scenarios, please refer to the corresponding or relevant descriptions in one or more embodiments of this specification, which will not be elaborated on here.

[0114] Sample images can be understood as CT images, plain CT images, enhanced CT images, pCT, and / or MRI images used as training samples. Sample labels can be understood as segmented images used to segment target objects in sample images; for example, the sample label can be a tumor segmentation image; the sample label can be a semantic segmentation mask. Object region information can be understood as information representing the location region of the target object in the sample image; for example, the object region information can be coordinates, a mask, or an anchor box. Flipped sample images can be understood as images obtained by flipping sample images, and the flipped sample labels can be understood as the sample labels corresponding to the flipped sample images.

[0115] Taking the application of the image processing model training method provided in this specification in the scenario of precise tumor segmentation as an example, the image processing model training method is explained. The sample image of the target object located at the target site is a CT image of a tumor located in the nasopharynx. The sample label is the semantic segmentation mask corresponding to the sample image. The object region information can be an asymmetric mask used to locate the tumor in the CT image. The flipped sample image can be a CT image obtained by flipping the CT image.

[0116] Based on this, this method determines the CT image of a tumor located in the nasopharynx, where the CT image I∈RD×H×W, and D, H, and W are the spatial dimensions of the CT image; and determines the flipped CT image I′ of CT image I.

[0117] A semantic segmentation mask is determined for CT image I, denoted as s∈{0,1}, with a shape of D×H×W, where 0 represents the background and 1 represents the tumor foreground. Based on this semantic segmentation mask s, a flipped semantic segmentation mask s′ for the flipped CT image I′ is determined, and s and s′ are used as the sample label and flipped sample label, respectively.

[0118] Then, an asymmetric mask m (i.e., object region information) is defined to locate the asymmetric region (i.e., the tumor region) in CT image I, i.e., m = ss ∩ s′, where m ∈ {0, 1} D × H × W. Note that 1 and 0 represent the asymmetric and symmetric regions in CT image I, respectively.

[0119] In one or more embodiments provided in this specification, in order to improve the accuracy of the image processing model, the training data can be augmented to obtain a larger amount of training data; the model can be trained using the larger amount of training data to improve the accuracy of the detection results output by the target image processing model; the specific implementation method is as follows.

[0120] The determination of the target training data and the augmented training data includes:

[0121] Determine the target training data;

[0122] The sample image is horizontally flipped to obtain the flipped sample image, and the sample label is horizontally flipped to obtain the flipped sample label of the flipped sample image;

[0123] The flipped sample image and the flipped sample label are used as the enhanced training samples.

[0124] Following the previous example, after determining the original CT image I, the CT image I is horizontally flipped to obtain the flipped CT image I′. After determining the semantic segmentation mask s of the CT image I, the flipped semantic mask s′ of the flipped CT image I′ can be obtained by performing a horizontal flipping operation on the semantic segmentation mask s.

[0125] In one or more embodiments provided in this specification, the target portion in the sample image and the flipped sample image provided by this method can be a part that satisfies semantic symmetry. Since the target part satisfies semantic symmetry, this method can accurately identify the target object in the target part based on semantic asymmetry. In order to improve the learning efficiency and effect of the image processing model, the sample image and the flipped sample image provided by this method include a central axis, which can symmetrically divide the target part. Specifically, the central axis is determined in the sample image and the flipped sample image as follows.

[0126] Determining the target training data includes:

[0127] Determine the object image of the target object located at the target location, and determine the object image to be processed from the object image, wherein the object image to be processed is any image in the object image;

[0128] Based on the target region, a central axis is added to the image of the object to be processed to obtain an initial object image, wherein the central axis symmetrically divides the target region;

[0129] Based on the initial object image, the central axis is added to other object images in the object image except for the object image to be processed, to obtain the target object image;

[0130] The initial object image and the target object image are used as the sample images, and the sample labels corresponding to the sample images and the object region information of the target object are determined.

[0131] The sample image, the sample label, and the object region information are used as the target training data.

[0132] The central axis can be the central sagittal axis or the sagittal axis.

[0133] The step of adding the central axis to other object images in the object image other than the object image to be processed, based on the initial object image, can be achieved by using the initial object image as a map, and then using affine registration to add the central axis to other object images in the object image other than the object image to be processed, thereby obtaining the target object image.

[0134] The method for determining the sample label corresponding to the sample image and the object region information of the target object can be found in one or more embodiments.

[0135] Following the previous example, the SATs (Image Processing Model) of this method focuses on asymmetrical abnormal regions in CT images. Considering that image asymmetry may originate from pathological or non-pathological sources, such as changes in imaging angle and patient posture, CT scan images can be preprocessed to ensure symmetry along the central sagittal axis. Specifically, this method selects bilaterally symmetrical patient CT images along the central sagittal plane as atlases. Then, based on these atlases, affine registration is used to align other patient CT images into the atlas space, thereby determining the central sagittal axis for other patient CT images. This step helps mitigate the influence of other asymmetrical anatomical structures in the head and neck that may mislead the model.

[0136] Step 204: Using the initial image processing model, perform object detection on the sample image and the flipped sample image respectively to obtain the detection results of the sample image and the flipped detection results of the flipped sample image.

[0137] The image processing model can be understood as a model that performs object detection processing on target objects in sample images or flipped sample images. This image processing model can be an image segmentation model or an object recognition model. The initial image processing model can be understood as the image processing model in any round during the iterative model training process. Correspondingly, the target image processing model can be the trained image processing model.

[0138] The detection result can be understood as the result obtained by the initial image processing model after performing object detection on the sample image. The detection result can be a tumor segmentation map, segmentation map, probability map, heat map, boundary map, or mask. The flip detection result can be understood as the detection result obtained by the initial image processing model after performing object detection on the flipped sample image.

[0139] Following the previous example, in the process of using an image segmentation model to identify tumors in CT images, the image segmentation model can assign each pixel in the input CT image to a specific category, thereby identifying tumor regions and non-tumor regions (background regions) in the CT. Based on this, after inputting CT image I and the flipped CT image I′ into the image segmentation model, the image segmentation model can assign each pixel in the input CT image to a specific category and output the segmentation maps corresponding to CT image I and the flipped CT image I′.

[0140] In one or more embodiments provided in this specification, the step of using an initial image processing model to perform object detection on the sample image and the flipped sample image respectively, to obtain the detection result of the sample image and the flipped detection result of the flipped sample image, includes:

[0141] Using the encoder in the initial image processing model, the sample image and the flipped sample image are encoded respectively to obtain the sample image encoding of the sample image and the flipped sample image encoding of the flipped sample image;

[0142] Using the decoder in the initial image processing model, the encoded sample image and the encoded flipped sample image are decoded respectively to obtain the detection result of the sample image and the flipped detection result of the flipped sample image.

[0143] Following the previous example, the SATs architecture provided by this method includes an encoder-decoder module and a projection head. The encoder-decoder is a U-shaped encoder-decoder architecture with shared weights, where the encoder employs repeated applications of 3D residual blocks, each block containing two convolutional layers with 3×3×3 kernels. Each convolutional layer is followed by InstanceNorm normalization and LeakyReLU activation. For downsampling, the resolution of the input feature map is halved using a convolution operation with a stride of 2. The initial number of filters is 32, which is doubled after each downsampling step to maintain constant time complexity, except for the last layer; in total, the encoder performs four downsampling operations.

[0144] Based on this, it can be seen that the SATs model of this method uses a Siamese segmentation network with shared weights for encoder-decoder to process the original CT image I and the flipped CT image I′.

[0145] It should be noted that although both the encoder-decoder module and the projector head are involved in the training process, the encoder-decoder module can be used during inference.

[0146] Step 206: Based on the detection results and the flip detection results, determine the predicted object region information of the target object in the sample image.

[0147] The predicted object region information can be understood as the object region information predicted using the detection results and the flip detection results. For example, the object region information can be tumor region feature encoding, tumor region coordinates, etc.

[0148] In one or more embodiments provided in this specification, the image processing model training method proposes a semantic asymmetric learning approach to achieve accurate GTV segmentation in nasopharyngeal carcinoma by leveraging the asymmetric characteristics of pharyngeal tumors. This image processing model uses an encoder-decoder with shared weights to process the original CT image I and the flipped CT image I′. Then, a nonlinear projection module and distance metric are used to learn and maximize the difference between abnormal and normal locations. Based on this, GTV segmentation in nasopharyngeal carcinoma can be significantly improved by utilizing the semantic symmetry in anatomical structures. This image processing model can be used in radiotherapy practice to standardize GTV depiction of nasopharyngeal carcinoma and reduce the workload and variability for radiation oncologists, as specifically implemented below.

[0149] The step of determining the predicted object region information of the target object in the sample image based on the detection results and the flip detection results includes:

[0150] The detection results are augmented using a nonlinear processing unit to obtain enhanced detection results, and the flip detection results are augmented using a nonlinear processing unit to obtain enhanced flip detection results.

[0151] Determine the object features corresponding to the target object included in the enhanced detection features, and determine the flipped object features of the target object included in the enhanced flip detection results;

[0152] The object features and the flipped object features are compared to determine the enhanced image features corresponding to the sample image;

[0153] From the enhanced image features, the predicted object region information of the target object in the sample image is determined.

[0154] The nonlinear processing unit can be understood as a nonlinear projection module, and the enhanced image feature can be understood as an enhanced image feature. The nonlinear processing unit can be a projection head.

[0155] Following the previous example, the SATs architecture provided by this method includes an encoder-decoder module and a projection head. This projection head, before calculating the distance of the marginal loss, uses a non-linear projection module to transform the features (i.e., the detection results and flipped detection results), aiming to improve the quality of the learned features. It consists of three 1×1×1 convolutional layers, each with 16 channels, followed by a unit normalization layer.

[0156] Specifically, the projection head consists of three convolutional layers (Conv), and the first two layers of the projection head can use the ReLU activation function; the specific execution method is as follows:

[0157] 1. Input the segmentation map of CT image I into the projection head (i.e., nonlinear processing unit), and use three convolutional layers to perform nonlinear enhancement processing on the segmentation map (i.e. detection result) of CT image I to obtain the CT enhancement image feature E (i.e. enhancement detection result) of CT image I.

[0158] 2. Input the segmentation map of the flipped CT image I′ (i.e., the flip detection result) into the projection head (i.e., the nonlinear processing unit), and use three convolutional layers to perform nonlinear enhancement processing on the segmentation map of the flipped CT image I′ to obtain the CT enhancement image feature Ef (i.e., the flipped object feature) of the flipped CT image I′.

[0159] 3. Tumor features identified from CT enhanced image feature E, and tumor features identified from CT enhanced image feature Ef;

[0160] 4. Compare the two tumor features to determine the CT image feature D (i.e., the enhanced image feature) corresponding to CT image I;

[0161] 5. Determine the tumor region feature code corresponding to the tumor from CT image feature D.

[0162] As can be seen from the above embodiments, this method assumes that directly applying metric learning to segmentation features may lead to information loss and reduce the effectiveness of the model. For example, some asymmetries in CT images are non-pathological and may stem from changes in the patient's head positioning and posture, but they are beneficial for segmentation. Using nonlinear projection can filter out this irrelevant information during the metric learning process, ensuring that it is preserved in the features used for segmentation.

[0163] In one or more embodiments provided in this specification, comparing the object features and the flipped object features to determine the enhanced image features corresponding to the sample image includes:

[0164] Using a nonlinear processing unit, it is determined that the object features contain multiple object feature points, and the flipped object features contain multiple flipped object feature points;

[0165] From the plurality of flipped object feature points, determine the corresponding feature points of each object feature point, and determine the similarity between each object feature point and the corresponding feature point;

[0166] If the similarity is greater than or equal to a first threshold, the similarity weight between the object feature point and the corresponding feature point is increased; if the similarity is greater than or equal to the first threshold, the similarity weight between the object feature point and the corresponding feature point is decreased, thereby obtaining the enhanced image features corresponding to the sample image.

[0167] Using the previous example, 1. From the pixels of the CT enhanced image feature Ef (i.e., the flipped object feature points), determine the pixels (i.e., the corresponding feature points) that correspond to the pixels of the CT enhanced image feature E (i.e., the object feature points); these pixels can be understood as features, feature codes, masks, and other information.

[0168] 2. Calculate the similarity between corresponding pixels in E and Ef to determine the distance between each pixel, and then obtain the pixel-by-pixel distance between E and Ef.

[0169] 3. Based on the pixel-by-pixel distance, the similarity between each pixel in E and each pixel in Ef is reduced or reduced to obtain the adjusted CT image feature D;

[0170] When the distance between pixels is greater than or equal to a preset similarity threshold, the similarity weight between two pixels is increased, thereby bringing the two pixels closer together.

[0171] When the distance between pixels is less than or equal to a preset dissimilarity threshold, the similarity weight between two pixels is reduced, thereby reducing the similarity between the two pixels.

[0172] As can be seen from the above embodiments, this method introduces a nonlinear projection module and a distance metric learning strategy into the SATs model to refine the generated feature map; thereby maximizing the difference between E and Ef at the corresponding anatomical locations of abnormality and normality.

[0173] Step 208: Based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label, train the initial image processing model to obtain the trained target image processing model.

[0174] In one or more embodiments provided in this specification, the target site is the nasopharyngeal region, the target object is the target tumor, the sample image is a sample CT image, and the flipped sample image is a flipped sample CT image.

[0175] The step of training the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flipped sample label to obtain the trained target image processing model includes:

[0176] Based on tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor results, and flipped sample labels, the initial image processing model is trained to obtain the trained target image processing model.

[0177] The explanations of the above embodiments can be referenced to the corresponding or relevant explanations in other embodiments, which will not be repeated here.

[0178] In one or more embodiments provided in this specification, training the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flipped sample label to obtain the trained target image processing model includes:

[0179] Based on the object region information and the predicted object region information, a first loss function is calculated, wherein the first loss function is used to evaluate the difference between the object region information and the predicted object region information.

[0180] Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a second loss function is calculated, wherein the second loss function is used to evaluate the difference between the detection results and the sample labels, and the difference between the flip detection results and the flipped sample labels;

[0181] Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a third loss function is calculated, wherein the third loss function is used to evaluate the similarity between the detection results and the sample labels, and the similarity between the flip detection results and the flipped sample labels;

[0182] Based on the first loss function, the second loss function, and the third loss function, the model parameters of the initial image processing model are adjusted to obtain the trained target image processing model.

[0183] Following the previous example, this method employs an asymmetric learning strategy. The segmentation loss function for the STA model consists of two parts: a combination of the Dice loss function (second loss function) used for traditional segmentation purposes and the cross-entropy loss function (second loss function), and a voxel marginal loss (i.e., first loss function) specifically designed for asymmetric outlier regions.

[0184] For the Dice loss function and the cross-entropy loss function, a normalization layer can be used to normalize the two segmentation maps (i.e., the detection result and the flipped detection result) to obtain the two normalized segmentation maps (i.e., the normalized detection result and the flipped detection result); then, the DSC loss and CE loss are calculated based on the two normalized segmentation maps, the sample labels, and the flipped sample labels.

[0185] For metric-based marginal loss, this method employs pixel-level marginal loss. In asymmetric anomaly regions, the goal is to minimize the similarity between the feature of any point and its corresponding point on the central sagittal axis; the marginal loss between feature E and the flipped feature E′ is:

[0186]

[0187] In the formula, the feature E∈RH×W×D×C, where H, W and D are spatial dimensions, C is the number of output features, 1 is an indicator function, t defines a boundary to adjust the degree of semantic asymmetry, (i, j, k) represents the position of the voxel (tumor region pixel), which can be understood as three-dimensional coordinates; m is a mask, representing a specific region.

[0188] This specification provides an image processing model training method through one or more embodiments. This method utilizes a large amount of target training data and augmented training data to train an initial image processing model, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on these detection and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Subsequently, based on this target image processing model, accurate detection results are obtained during object detection of images.

[0189] The following is in conjunction with the appendix Figure 4 Taking the application of the image processing model training method provided in this specification in the nasopharyngeal carcinoma scenario as an example, the image processing model training method will be further explained. Among other things, Figure 4 This specification illustrates a flowchart of the processing procedure for an image processing model training method according to an embodiment of this specification. Figure 4 It can be seen that the SATs architecture provided by this method includes an encoder-decoder module and a projection head;

[0190] The encoder-decoder module employs a U-shaped encoder-decoder architecture with shared weights. The encoder uses repeated applications of 3D residual blocks, each containing two convolutional layers with 3×3×3 kernels. Each convolutional layer is followed by InstanceNorm normalization and LeakyReLU activation.

[0191] For downsampling, a convolution operation with a stride of 2 is used to halve the resolution of the input feature map; the initial number of filters is 32, which is doubled after each downsampling step to maintain constant time complexity, except for the last layer. In total, the encoder performs four downsampling operations.

[0192] Here, the projection head also refers to nonlinear projection. Before calculating the distance of the marginal loss, this scheme uses a nonlinear projection module to transform the image features output by the encoder-decoder module (i.e., the model output of the SATs architecture) in order to improve the quality of the learned features.

[0193] The projection head consists of three 1×1×1 convolutional layers, each with 16 channels, followed by a unit normalized layer. The first two layers of the projection head use the ReLU activation function. This scheme assumes that directly applying metric learning to segmentation features may lead to information loss and reduce model effectiveness. For example, some asymmetries in CT images are non-pathological and may stem from changes in patient head positioning and pose, but they are beneficial for segmentation. Therefore, non-linear projection can be used to filter out this irrelevant information from the metric learning process, ensuring that it is preserved in the features used for segmentation.

[0194] It should be noted that the following steps are based on the application of the image processing model training method provided in this manual in the nasopharyngeal carcinoma scenario, to explain the image processing model training method. The image processing model training method provided in this manual can be applied to cancer scenarios for computer-aided diagnosis of cancer, to breast cancer scenarios for computer-aided diagnosis of breast cancer, or to pelvic cancer scenarios for computer-aided diagnosis of pelvic cancer. For the specific execution methods for cancer, breast cancer, and pelvic cancer scenarios, please refer to the corresponding or relevant descriptions in the following steps, which will not be elaborated on here.

[0195] Based on this, the image processing model training method specifically includes the following steps.

[0196] Step 402: Construct training data.

[0197] The training data for this scheme includes original CT images and flipped CT images. The specific method for determining the training data is as follows.

[0198] 1. Obtain CT image I as the sample image.

[0199] Here, CT image I refers to the CT image of a tumor located in the nasopharynx, and this CT image I ∈ RD×H×W, where D, H, and W are the spatial dimensions of the CT image.

[0200] It should be noted that the CT image (i.e. the CT image to be detected in the embodiments of this specification) may be a CT image to be detected of a target tumor located in the target area, a CT image to be detected of a target tumor located in the nasopharynx, a CT image to be detected of a target tumor located in the breast area, or a CT image to be detected of a target tumor located in the pelvic area.

[0201] 2. Select the corresponding CT images along both sides of the central sagittal plane from CT image I as the atlas;

[0202] 3. Based on this atlas, affine registration is used to align other patients' CT images to the atlas space, thereby determining the central sagittal axis for other patients' CT images.

[0203] This step helps mitigate the influence of other asymmetrical anatomical structures in the head and neck that may mislead the model.

[0204] 4. Mark the lesion areas in the CT images to obtain sample labels.

[0205] Specifically, this scheme can determine a semantic segmentation mask for CT image I. The semantic segmentation mask is represented as s∈{0,1}, and the shape of the semantic segmentation mask is D×H×W, where 0 represents the background and 1 represents the tumor foreground.

[0206] 5. Flip the CT image I to obtain the flipped CT image I′. Figure 4 I in f ) and the corresponding sample labels.

[0207] Specifically, after determining the original CT image I, the CT image I is horizontally flipped to obtain the flipped CT image I′. After determining the semantic segmentation mask s of the CT image I, the flipped semantic mask s′ of the flipped CT image I′ can be obtained by performing a horizontal flipping operation on the semantic segmentation mask s.

[0208] 6. Mark the asymmetric regions in CT image I to obtain the asymmetric mask.

[0209] Specifically, this method defines an asymmetric mask m to locate the asymmetric region (i.e., the tumor region) in CT image I, i.e., m = ss∩s′, where m∈{0,1}D×H×W. It should be noted that 1 and 0 represent the asymmetric and symmetric regions in CT image I, respectively.

[0210] Step 404: Input the training samples into the SATs model to obtain the segmentation map output by the SATs model.

[0211] Specifically, the SATs model in this scheme includes an encoder and a decoder. The encoder and decoder are used to encode and decode CT image I and the flipped CT image I′ respectively to obtain the segmentation maps corresponding to CT image I and the flipped CT image I′. The specific implementation method is as follows:

[0212] In the process of using image segmentation models to identify tumors in CT images, the image segmentation model can assign each pixel in the input CT image to a specific category, thereby identifying tumor areas and non-tumor areas (background areas) in the CT.

[0213] Based on this, after inputting CT image I and flipped CT image I′ into the image segmentation model, the image segmentation model can use an encoder to encode CT image I and flipped CT image I′, thereby obtaining the image codes corresponding to CT image I and flipped CT image I′.

[0214] Then, the image segmentation model can use a decoder to decode the image codes corresponding to CT image I and flipped CT image I′, thereby assigning each pixel in CT image I and flipped CT image I′ to a specific category and outputting the segmentation map corresponding to CT image I and flipped CT image I′.

[0215] The specific category can be (0, 1), where 0 represents non-tumor pixels and 1 represents tumor pixels; the segmentation map can be a semantic segmentation mask composed of (0, 1).

[0216] Step 406: Normalize the segmentation map using a normalization layer and calculate DSC loss and CE Ioss.

[0217] Step 408: Utilize the CT enhanced image features E and Ef obtained from the projection head; based on the metric learning strategy, obtain the adjusted CT image features D, and calculate the marginal loss function based on the adjusted CT image features D and the asymmetric mask m.

[0218] Specifically, the projection head consists of three convolutional layers (Conv), and the specific execution method is as follows:

[0219] 1. Input the segmentation map of CT image I into the projection head (i.e., the nonlinear projection module), and use three convolutional layers to perform nonlinear enhancement processing on the segmentation map of CT image I to obtain the CT enhancement image feature E of CT image I;

[0220] 2. Input the segmented image of the flipped CT image I′ into the projection head (i.e., the nonlinear projection module), and use three convolutional layers to perform nonlinear enhancement processing on the segmented image of the flipped CT image I′ to obtain the CT enhancement image features Ef of the flipped CT image I′. Figure 4 E in f );

[0221] 3. From the pixels of the CT enhanced image feature Ef, determine the pixels corresponding to the pixels of the CT enhanced image feature E, and calculate the similarity between the corresponding pixels in E and Ef to determine the distance between each pixel, and then obtain the pixel-by-pixel distance between E and Ef.

[0222] 4. Based on the pixel-by-pixel distance, the similarity between each pixel in E and each pixel in Ef is reduced or reduced to obtain the adjusted CT image features D.

[0223] When the distance between pixels is greater than or equal to a preset similarity threshold, the similarity weight between two pixels is increased, thereby bringing the two pixels closer together.

[0224] When the distance between pixels is less than or equal to a preset dissimilarity threshold, the similarity weight between two pixels is reduced, thereby reducing the similarity between the two pixels.

[0225] 5. Calculate MarginLoss based on the tumor region feature encoding and asymmetric mask m in the CT image feature D.

[0226] It should be noted that after obtaining the Margin Loss, DSC loss, and CE loss, the three loss functions are added together to obtain the overall loss function. This overall loss function is then used to adjust the model parameters of the SATs model until the model training stops, thus obtaining the trained SATs model.

[0227] Based on the above steps, the image processing model training method in this specification proposes a precise tumor segmentation method for nasopharyngeal carcinoma in non-contrast CT images using semantic asymmetry. This method is based on the 3D semantic asymmetric tumor segmentation (SATs) method, which utilizes the semantic asymmetry characteristics of gross tumors in the nasopharyngeal region. It can achieve accurate NPC GTV segmentation even with limited data. This method can directly segment NPC GTVs in non-contrast pCT, thus avoiding registration errors that occur when aligning MRI or MRI-derived tumor masks with pCT. It also solves the problem of poor contrast between GTVs and adjacent normal structures in pCT.

[0228] In this method, the goal is to segment NPC GTV in non-contrast pCT (the most common RT imaging), which avoids registration errors when aligning MRI or MRI-derived tumor masks with pCT. Direct segmentation of NPC GTV in non-contrast pCT is challenging because the boundaries between the NPC tumor and adjacent soft tissues (such as membranes, muscles, and blood vessels) are very unclear in non-contrast pCT. To alleviate this problem, this method proposes a 3D semantic asymmetric tumor segmentation (SATs) approach based on the observation that healthy nasopharyngeal regions are typically bilaterally symmetrical (see...). Figure 3 However, the presence of NPC tumors disrupts this symmetry. Specifically, to leverage anatomical symmetry cues, the pCT is first wrapped with automatically segmented head and neck organs to ensure bilateral symmetry along the central sagittal plane. This helps reduce the impact of asymmetrical anatomy caused by changes in patient head position and posture during CT scanning. Then, a Siamese contrastive learning segmentation framework is developed based on conventional segmentation losses (Dice and / or cross-entropy losses) and additional voxel-level edge losses. The edge loss is applied to depth features extracted from the original pCT scan and the flipped pCT scan, aiming to minimize the voxel distance between the original and flipped tumor-free regions and encourage a larger distance between the original and flipped tumor-containing regions, making GTV features more sensitive to semantic asymmetry.

[0229] Based on the above, it is evident that the method described in this specification can effectively improve the performance of NPC GTV segmentation in non-contrast pCT by utilizing the inherent semantic symmetry of anatomical structures.

[0230] During training, this method employs the stochastic gradient descent algorithm with a momentum of 0.99 as the optimizer. Model training is divided into two phases. In the first phase, only the Siamese encoder is trained 800 times with a learning rate of 1e-2, decaying using polynomial scheduling. In the second phase, the projection head is trained 200 times with a learning rate of 1e-2, and the Siamese encoder has a learning rate of 1e-5, both decaying using polynomial scheduling. The patch size is 56×192×192, and the batch size is 2. A sliding window strategy is used for testing. For voxel contrast loss, this method uses an edge hyperparameter t=20.

[0231] During data preparation, an internal dataset was collected and organized from the hospital for the development of segmentation models (SATs). This dataset included pCT scans, contrast-enhanced diagnostic CT scans, and diagnostic MRI scans of T1 and T2 stages from 163 nasopharyngeal carcinoma patients. The diagnostic CT and MRI scans were registered. First, the MRI images underwent rigid transformation to approximately align with the CT images, ensuring similar anatomical localization. Subsequently, precise local alignment was achieved using the cross-modal deformable registration algorithm DEEDs. Contrast-enhanced CT and MRI scans were used to guide radiation oncologists in generating ground-truth (GTV) values ​​in pCT scans. Additionally, this method collected and organized a publicly available dataset as an external dataset, containing 118 non-contrast pCT and contrast-enhanced CT scans. The original tumor targets in publicly available datasets are often larger than described in GTV guidelines. Therefore, their labels were organized. Annotations for all datasets were reviewed and edited by two experienced radiation oncologists according to international consensus guidelines for GTV delineation. For evaluation, 20% of the internal dataset was randomly selected as the internal test set, and a curated publicly available dataset was used as the external test dataset. The asymmetric regions in the external data were larger than those in the internal data, making the segmentation task more challenging.

[0232] It should be noted that regarding comparison methods and evaluation metrics: our method is comprehensively compared with various state-of-the-art methods, including CNN-based methods, transformer-based methods, and Mambaba-based methods (a medical image segmentation method), to evaluate its performance. To ensure a fair comparison, all competing models were trained for the same number of epochs, 1000. To avoid model crashes during training, different learning rates were used for different comparison methods. Generally, CNN-based methods use larger learning rates, while transformer-based methods use smaller learning rates. For example, for various state-of-the-art methods, the learning rate was set to 1e-2, and stochastic gradient descent with momentum of 0.99 was used as the optimizer. For TransUNet (a CNN-based method), the learning rate was set to 1e-3, and stochastic gradient descent with momentum of 0.99 was used as the optimizer. For UNETR (a transformer-based method), the learning rate was set to 1e-5, and the AdamW optimizer was used. Furthermore, this method uses the Dice similarity coefficient (DSC%) to evaluate the overlap of segmentation maps and calculates the average surface distance (ASD) (mm) for all cases. Additionally, this method measures the 95th percentile (HD95) of the Hausdorff distance (mm) to measure registration accuracy relative to the structural boundary. Generally, a better method should produce a higher DSC score and lower ASD and HD95 values.

[0233] In comparing our approach with state-of-the-art methods, to validate GTV segmentation of NPCs in non-contrast CT images, we compared our proposed SATs model with existing state-of-the-art (SOTA) segmentation methods using an internal dataset. With relatively few parameters, our proposed SATs model significantly improves upon existing SOTA methods. Specifically, compared to suboptimal methods, our proposed SATs model improves DSC, ASD, and HD95 by 1.0%, 2.6%, and 3.6%, respectively. Our proposed SATs model also demonstrates higher accuracy in boundary segmentation (e.g., nasal septum).

[0234] To evaluate the cross-domain transfer learning, two experiments were conducted to verify the generalization performance of our method across different domains: First, we trained our method using the training set of an internal dataset and validated it on the test set of an external dataset; second, we trained our method using the training set of an external dataset and validated it on the test set of an internal dataset. Our method demonstrated excellent performance in both experiments, particularly in the first experiment, thus confirming the effectiveness of the learned features related to asymmetric tumors. Notably, our method achieved the highest mean and lowest variability in both demanding tasks, demonstrating superior and consistent performance. Furthermore, it is evident that our method has a significant advantage when the internally trained model is applied to external data with more pronounced asymmetric regions.

[0235] In conducting ablation studies, the effects of the projection head (Proj.Head) and marginal loss (Marg.Loss) on the segmentation model were investigated. Table 1 shows the effects on external data. Performance metrics of different segmentation models.

[0236] Table 1

[0237]

[0238] As shown in Table 1, although margin loss was added, the model still had a lower DSC and a higher ASD / HD95 compared to the initial model, indicating that margin loss alone was insufficient to improve performance. Furthermore, when the projection head module and margin loss were added to the baseline, there was a significant performance improvement (+4.98% DSC, -0.60mm ASD, and -1.15mm HD95). Table 1 also shows that semantic asymmetric learning, consisting of the projection head and margin loss, effectively improved the accuracy of tumor segmentation; this relates to the impact of semantic asymmetric learning. The proposed method was compared with a model without a projection head module and / or a baseline with margin loss. The proposed method achieved a higher Dice score and consistently outperformed other models in most tests across 117 test scans.

[0239] It should be noted that the embodiments in this specification are for semantic asymmetric learning using a large amount of unlabeled data. This method can use the proposed unsupervised asymmetric region selection method (through inverse consistency constraints of bidirectional deformable image registration) to collect additional unlabeled data and detect asymmetric anomalous regions. For unlabeled data, this method calculates the marginal loss during training. This method can capture useful GTV segmentation representations of asymmetric regions in unlabeled data, achieving improved performance compared to the benchmark SAT; based on this, the method proposed in this specification has the potential to address the scarcity of labeled data in the field of medical image analysis.

[0240] In summary, this specification presents one or more embodiments of a semantic asymmetric learning method to achieve accurate GTV segmentation of nasopharyngeal carcinoma (NPC) by leveraging the asymmetric characteristics of tumors in the pharyngeal region. The method uses a Siamese segmentation network with a shared encoder-decoder to process raw and flipped CT images, followed by a nonlinear projection module and distance metric learning to maximize the differentiation between abnormal and normal locations. By utilizing the semantic symmetry of anatomical structures, this method significantly improves GTV segmentation for NPC, outperforming nine state-of-the-art methods evaluated on an internal test set and an independent external dataset. In radiotherapy practice, standardizing the delineation of NPC GTV reduces the workload and variability for radiation oncologists and has potential applications.

[0241] See Figure 5 , Figure 5 A flowchart of another image processing model training method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0242] Step 502: Determine the target training data and the augmented training data, wherein the target training data includes a sample CT image of a target tumor located in the nasopharynx, a sample label of the sample CT image, and tumor region information of the target tumor; the augmented training data includes a flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image.

[0243] Step 504: Using the initial image processing model, perform tumor detection on the sample CT image and the flipped sample CT image respectively to obtain the tumor detection results of the sample CT image and the flipped tumor detection results of the flipped sample CT image:

[0244] Step 506: Based on the tumor detection results and the flipped tumor detection results, determine the predicted tumor region information of the target tumor in the sample CT image;

[0245] Step 508: Based on the tumor region information, the predicted tumor region information, the tumor detection result, the sample label, the flipped tumor detection result, and the flipped sample label, train the initial image processing model to obtain the trained target image processing model.

[0246] This specification provides another image processing model training method through one or more embodiments. This method utilizes a large amount of target training data and augmented training data to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve tumor detection accuracy, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined. Then, the tumor region information and predicted tumor region information are used to train the initial image processing model, thereby achieving accurate detection of the target tumor by the target image processing model. Therefore, by training the initial image processing model using tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a target image processing model capable of more accurate tumor detection is obtained. Subsequently, based on this target image processing model, accurate tumor detection results can be obtained during tumor detection of CT images.

[0247] The above is an illustrative scheme of another image processing model training method in this embodiment. It should be noted that the technical solution of this other image processing model training method belongs to the same concept as the technical solution of the image processing model training method described above. For details not described in detail in the technical solution of the other image processing model training method, please refer to the description of the technical solution of the image processing model training method described above.

[0248] See Figure 6 , Figure 6 A flowchart of an image processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0249] Step 602: Determine the image to be detected of the target object located at the target location;

[0250] Step 604: Perform object detection on the image to be detected using a target image processing model to obtain the detection result of the image to be detected. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection result, sample label, flip detection result, and flip sample label. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by performing object detection on the sample image using the initial image processing model. The flip detection result is determined by performing object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0251] This specification provides an image processing method in one or more embodiments. To achieve accurate detection results using a target image processing model, an initial image processing model can be trained using a large amount of target training data and augmented training data, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on the detection results and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Then, object detection is performed on the image to be detected based on this target image processing model, thereby obtaining accurate detection results.

[0252] The above is an illustrative scheme of an image processing method according to this embodiment. It should be noted that the technical solution of this image processing method belongs to the same concept as the technical solution of the image processing model training method described above. For details not described in detail in the technical solution of the image processing method, please refer to the description of the technical solution of the image processing model training method described above.

[0253] See Figure 7, Figure 7 A flowchart of another image processing method according to an embodiment of this specification is shown. This other image processing method is applied to a client of a medical system and specifically includes the following steps.

[0254] Step 702: In response to the user's selection operation on the client's display interface, determine the medical image to be detected;

[0255] Step 704: Send the medical image to be detected to the server of the medical system and receive the detection result returned by the server. The detection result is the output result after object detection of the medical image to be detected by the target image processing model. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection result, sample label, flip detection result, and flip sample label. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by object detection of the sample image using the initial image processing model. The flip detection result is determined by object detection of the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0256] Step 706: Display the detection results to the user through the display interface.

[0257] Another image processing method provided in one or more embodiments of this specification aims to obtain accurate detection results using a target image processing model. This method utilizes a large amount of target training data and augmented training data to train an initial image processing model, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on these detection and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Finally, based on this target image processing model, object detection is performed on the image to be detected sent by the client, thereby obtaining accurate detection results, which are then returned to the client for display.

[0258] The above is an illustrative scheme of another image processing method in this embodiment. It should be noted that the technical solution of this other image processing method belongs to the same concept as the technical solution of the image processing model training method described above. For details not described in detail in the technical solution of the other image processing method, please refer to the description of the technical solution of the image processing model training method described above.

[0259] See Figure 8 , Figure 8 A flowchart of an information processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0260] Step 802: Identify the CT image to be examined for the target tumor located in the nasopharynx:

[0261] Step 804: Use a target image processing model to perform tumor detection on the CT image to be detected, and obtain the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by performing tumor detection on the sample CT image using the initial image processing model. The flipped tumor detection result is determined by performing tumor detection on the flipped sample CT image using the initial image processing model. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0262] Step 806: Determine the information processing result based on the tumor detection result of the CT image to be detected.

[0263] The information processing result can be understood as the result obtained by processing the information based on the tumor detection result. For example, the information processing result can be the regional information of the tumor, the tumor type, the radiotherapy strategy for the tumor, etc., without being specifically limited here.

[0264] This specification provides an information processing method in one or more embodiments. During information processing, a large amount of target training data and enhanced training data can be used to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, to further improve the accuracy of tumor detection during model training, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined. Then, the tumor region information and predicted tumor region information are used to train the initial image processing model, thereby achieving accurate detection of the target tumor by the target image processing model. Therefore, by training the initial image processing model using tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a target image processing model capable of more accurate tumor detection is obtained. Subsequently, information processing is performed on the accurate tumor detection results output by the target image processing model to obtain accurate information processing results.

[0265] The above is an illustrative scheme of an information processing method according to this embodiment. It should be noted that the technical solution of this information processing method and the technical solution of the image processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the information processing method, please refer to the description of the technical solution of the image processing model training method described above.

[0266] This specification provides one or more embodiments of a computer-aided cancer diagnosis method, comprising:

[0267] Identify the CT image of the target tumor located at the target site;

[0268] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located at the target site, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0269] This specification provides one or more embodiments of a cancer computer-aided diagnosis method. During the cancer computer-aided diagnosis process, a large amount of target training data and augmented training data can be used to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve the accuracy of tumor detection, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the accuracy of tumor detection is determined... The system first obtains the predicted tumor region information of the target tumor in a sample CT image. Then, it uses the tumor region information and the predicted tumor region information to train the initial image processing model, thereby achieving accurate detection of the target tumor. Therefore, by training the initial image processing model using tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a more accurate target image processing model for tumor detection can be obtained. Subsequently, this target image processing model can be used to perform tumor detection on the CT image of the target tumor located in the target area, thereby outputting accurate tumor detection results for cancer.

[0270] The above is an illustrative scheme of a computer-aided cancer diagnosis method according to this embodiment. It should be noted that the technical solution of this computer-aided cancer diagnosis method and the technical solution of the image processing model training method described above belong to the same concept. Details not described in detail in the technical solution of the computer-aided cancer diagnosis method can be found in the description of the technical solution of the image processing model training method described above.

[0271] This specification provides an embodiment of a computer-aided diagnostic method for nasopharyngeal carcinoma, comprising:

[0272] Identify the CT image of the target tumor located in the nasopharynx:

[0273] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0274] This specification provides one or more embodiments of a computer-aided diagnostic method for nasopharyngeal carcinoma. During the computer-aided diagnosis of nasopharyngeal carcinoma, a large amount of target training data and enhanced training data can be used to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve the accuracy of tumor detection, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the method can determine the accuracy of tumor detection. The system first obtains the predicted tumor region information of the target tumor in a sample CT image. Then, it uses the tumor region information and the predicted tumor region information to train the initial image processing model, thereby achieving accurate detection of the target tumor. Therefore, by training the initial image processing model with tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a more accurate target image processing model for tumor detection can be obtained. Subsequently, this target image processing model can be used to detect tumors in CT images of target tumors located in the nasopharynx, thereby outputting accurate tumor detection results for nasopharyngeal carcinoma.

[0275] The above is an illustrative scheme of a computer-aided diagnosis method for nasopharyngeal carcinoma according to this embodiment. It should be noted that the technical solution of this computer-aided diagnosis method for nasopharyngeal carcinoma belongs to the same concept as the technical solution of the image processing model training method described above. Details not described in detail in the original computer-aided diagnosis method for nasopharyngeal carcinoma can be found in the description of the technical solution of the computer-aided diagnosis model training method for nasopharyngeal carcinoma described above.

[0276] This specification provides an embodiment of a computer-aided diagnosis method for breast cancer, comprising:

[0277] Identify the CT image of the target tumor located in the breast region;

[0278] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the breast region, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0279] This specification provides one or more embodiments of a computer-aided diagnosis method for breast cancer. During the computer-aided diagnosis of breast cancer, a large amount of target training data and augmented training data can be used to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve the accuracy of tumor detection, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the accuracy of the diagnosis can be determined. The system first obtains the predicted tumor region information of the target tumor in a sample CT image. Then, it uses the tumor region information and the predicted tumor region information to train the initial image processing model, thereby achieving accurate detection of the target tumor. Therefore, by training the initial image processing model with tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a more accurate target image processing model for tumor detection can be obtained. Subsequently, this target image processing model can be used to detect tumors in CT images of target tumors located in the breast region, thereby outputting accurate tumor detection results for breast cancer.

[0280] The above is an illustrative scheme of a computer-aided diagnosis method for breast cancer according to this embodiment. It should be noted that the technical solution of this computer-aided diagnosis method for breast cancer and the technical solution of the image processing model training method described above belong to the same concept. Details not described in detail in the technical solution of the computer-aided diagnosis method for breast cancer can be found in the description of the technical solution of the image processing model training method described above.

[0281] This specification provides an embodiment of a computer-aided diagnostic method for pelvic cancer, comprising:

[0282] Identify the CT image of the target tumor located in the pelvic region;

[0283] A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the pelvic region, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0284] This specification provides one or more embodiments of a computer-aided diagnosis method for pelvic cancer. During the computer-aided diagnosis of pelvic cancer, a large amount of target training data and enhanced training data can be used to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve the accuracy of tumor detection, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the accuracy of tumor detection is determined... The system first obtains the predicted tumor region information of the target tumor in a sample CT image. Then, it uses the tumor region information and the predicted tumor region information to train the initial image processing model, thereby achieving accurate detection of the target tumor. Therefore, by training the initial image processing model with tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a more accurate target image processing model for tumor detection can be obtained. Subsequently, this target image processing model can be used to detect tumors in CT images of target tumors located in the breast region, thereby outputting accurate tumor detection results for pelvic cancer.

[0285] The above is an illustrative scheme of a computer-aided diagnosis method for pelvic cancer according to this embodiment. It should be noted that the technical solution of this computer-aided diagnosis method for pelvic cancer belongs to the same concept as the technical solution of the image processing model training method described above. Details not described in detail in the technical solution of the computer-aided diagnosis method for pelvic cancer can be found in the description of the technical solution of the image processing model training method described above.

[0286] Corresponding to the above method embodiments, this specification also provides an embodiment of an image processing model training device. Figure 9 A schematic diagram of an image processing model training apparatus according to one embodiment of this specification is shown. Figure 9 As shown, the device includes:

[0287] The training data determination module 902 is configured to determine target training data and augmentation training data, wherein the target training data includes a sample image of a target object located at a target location, a sample label of the sample image, and object region information of the target object, and the augmentation training data includes a flipped sample image corresponding to the sample image and a flipped sample label of the flipped sample image.

[0288] The object detection module 904 is configured to use an initial image processing model to perform object detection on the sample image and the flipped sample image respectively, and obtain the detection result of the sample image and the flipping detection result of the flipped sample image.

[0289] The region information determination module 906 is configured to determine the predicted object region information of the target object in the sample image based on the detection result and the flip detection result;

[0290] The model training module 908 is configured to train the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label, to obtain the trained target image processing model.

[0291] Optionally, the model training module 908 is further configured to:

[0292] Based on the object region information and the predicted object region information, a first loss function is calculated, wherein the first loss function is used to evaluate the difference between the object region information and the predicted object region information.

[0293] Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a second loss function is calculated, wherein the second loss function is used to evaluate the difference between the detection results and the sample labels, and the difference between the flip detection results and the flipped sample labels;

[0294] Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a third loss function is calculated, wherein the third loss function is used to evaluate the similarity between the detection results and the sample labels, and the similarity between the flip detection results and the flipped sample labels;

[0295] Based on the first loss function, the second loss function, and the third loss function, the model parameters of the initial image processing model are adjusted to obtain the trained target image processing model.

[0296] Optionally, the area information determination module 906 is further configured to:

[0297] The detection results are augmented using a nonlinear processing unit to obtain enhanced detection results, and the flip detection results are augmented using a nonlinear processing unit to obtain enhanced flip detection results.

[0298] Determine the object region feature information corresponding to the target object contained in the enhanced detection features, and determine the flipped object region feature information of the target object contained in the enhanced flip detection result;

[0299] By comparing the object region feature information and the flipped object region feature information, the predicted object region information of the target object in the sample image is determined.

[0300] Optionally, the area information determination module 906 is further configured to:

[0301] Using a nonlinear processing unit, the object features of the target object in the detection result and the flipped object features of the target object in the flip detection result are determined, wherein the object features include multiple object feature points and the flipped object features include multiple flipped object feature points;

[0302] From the plurality of flipped object feature points, determine the corresponding feature points of each object feature point, and determine the similarity between each object feature point and the corresponding feature point;

[0303] If the similarity is greater than or equal to a first threshold, the similarity weight between the object feature point and the corresponding feature point is increased; if the similarity is greater than or equal to the first threshold, the similarity weight between the object feature point and the corresponding feature point is decreased, thereby obtaining the enhanced detection result and the enhanced flip detection result.

[0304] Optionally, the training data determination module 902 is further configured to:

[0305] Determine the target training data:

[0306] The sample image is horizontally flipped to obtain the flipped sample image, and the sample label is horizontally flipped to obtain the flipped sample label of the flipped sample image;

[0307] The flipped sample image and the flipped sample label are used as the enhanced training samples.

[0308] Optionally, the object detection module 904 is further configured to:

[0309] Using the encoder in the initial image processing model, the sample image and the flipped sample image are encoded respectively to obtain the sample image encoding of the sample image and the flipped sample image encoding of the flipped sample image;

[0310] Using the decoder in the initial image processing model, the encoded sample image and the encoded flipped sample image are decoded respectively to obtain the detection result of the sample image and the flipped detection result of the flipped sample image.

[0311] Optionally, the training data determination module 902 is further configured to:

[0312] Determine the object image of the target object located at the target location, and determine the object image to be processed from the object image, wherein the object image to be processed is any image in the object image;

[0313] Based on the target region, a central axis is added to the image of the object to be processed to obtain an initial object image, wherein the central axis symmetrically divides the target region:

[0314] Based on the initial object image, the central axis is added to other object images in the object image except for the object image to be processed, to obtain the target object image;

[0315] The initial object image and the target object image are used as the sample images, and the sample labels corresponding to the sample images and the object region information of the target object are determined.

[0316] The sample image, the sample label, and the object region information are used as the target training data.

[0317] Optionally, the target location is the nasopharyngeal region, the target object is the target tumor, the sample image is a sample CT image, and the flipped sample image is a flipped sample CT image;

[0318] The model training module 908 is further configured as follows:

[0319] Based on tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor results, and flipped sample labels, the initial image processing model is trained to obtain the trained target image processing model.

[0320] This specification provides an image processing model training apparatus in one or more embodiments. This apparatus utilizes a large amount of target training data and augmented training data to train an initial image processing model, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on the detection results and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Subsequently, based on this target image processing model, accurate detection results are obtained during object detection of images.

[0321] The above is a schematic scheme of an image processing model training device according to this embodiment. It should be noted that the technical solution of this image processing model training device and the technical solution of the image processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the image processing model training device, please refer to the description of the technical solution of the image processing model training method described above.

[0322] Corresponding to the above method embodiments, this specification also provides another embodiment of an image processing model training device. Figure 10 A schematic diagram of another image processing model training apparatus provided in one embodiment of this specification is shown. Figure 10 As shown, the device includes:

[0323] The training data determination module 1002 is configured to determine target training data and augmentation training data. The target training data includes a sample CT image of a target tumor located in the nasopharynx, a sample label of the sample CT image, and tumor region information of the target tumor. The augmentation training data includes a flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image.

[0324] The tumor detection module 1004 is configured to use an initial image processing model to perform tumor detection on the sample CT image and the flipped sample CT image respectively, and obtain the tumor detection result of the sample CT image and the flipped tumor detection result of the flipped sample CT image.

[0325] The region information determination module 1006 is configured to determine the predicted tumor region information of the target tumor in the sample CT image based on the tumor detection results and the flipped tumor detection results.

[0326] The model training module 1008 is configured to train the initial image processing model based on the tumor region information, the predicted tumor region information, the tumor detection result, the sample label, the flipped tumor detection result, and the flipped sample label, to obtain the trained target image processing model.

[0327] This specification provides another image processing model training apparatus, which can train an initial image processing model using a large amount of target training data and enhanced training data, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve the accuracy of tumor detection, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined. Then, the tumor region information and predicted tumor region information are used to train the initial image processing model, thereby achieving accurate detection of the target tumor by the target image processing model. Therefore, by training the initial image processing model using tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a target image processing model capable of more accurate tumor detection is obtained. Subsequently, based on this target image processing model, accurate tumor detection results can be obtained during tumor detection of CT images.

[0328] The above is an illustrative scheme of another image processing model training device according to this embodiment. It should be noted that the technical solution of this other image processing model training device and the technical solution of the other image processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the other image processing model training device, please refer to the description of the technical solution of the other image processing model training method described above.

[0329] Corresponding to the above method embodiments, this specification also provides an embodiment of an image processing apparatus. Figure 11 A schematic diagram of the structure of an image processing apparatus provided in one embodiment of this specification is shown. Figure 11 As shown, the device includes:

[0330] Image determination module 1102 is configured to determine the image to be detected of a target object located at the target location;

[0331] The object detection module 1104 is configured to perform object detection on the image to be detected using a target image processing model to obtain the detection result of the image to be detected. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by performing object detection on the sample image using the initial image processing model, and the flip detection result is determined by performing object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0332] This specification provides an image processing apparatus in one or more embodiments. To achieve accurate detection results using a target image processing model, an initial image processing model can be trained using a large amount of target training data and augmented training data, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve object detection accuracy, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on the detection results and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Then, object detection is performed on the image to be detected based on this target image processing model, thereby obtaining accurate detection results.

[0333] The above is an illustrative scheme of an image processing apparatus according to this embodiment. It should be noted that the technical solution of this image processing apparatus and the technical solution of the image processing method described above belong to the same concept. For details not described in detail in the technical solution of the image processing apparatus, please refer to the description of the technical solution of the image processing method described above.

[0334] Corresponding to the above method embodiments, this specification also provides another embodiment of an image processing apparatus. Figure 12 A schematic diagram of another image processing apparatus provided in one embodiment of this specification is shown. Figure 12 As shown, this other image processing device is applied to a client of a medical system, and the device includes:

[0335] The image determination module 1202 is configured to determine the medical image to be detected in response to a user's selection operation on the display interface of the client.

[0336] The object detection module 1204 is configured to send the medical image to be detected to the server of the medical system and receive the detection result returned by the server. The detection result is the output result after object detection is performed on the medical image by a target image processing model. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by object detection of the sample image using the initial image processing model. The flip detection result is determined by object detection of the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result.

[0337] The result display module 1206 is configured to display the detection results to the user through the display interface.

[0338] Another image processing apparatus provided in one or more embodiments of this specification, in order to obtain accurate detection results using a target image processing model, can train an initial image processing model using a large amount of target training data and augmented training data, thereby improving the accuracy of the detection results output by the target image processing model. Furthermore, during model training, to further improve the accuracy of object detection, the initial image processing model can be used to perform object detection on both sample images and flipped sample images, obtaining detection results for the sample images and flipped detection results for the flipped sample images. Based on the detection results and flipped detection results, predicted object region information of the target object in the sample image is determined. Then, the object region information and predicted object region information are used to train the initial image processing model, thereby achieving accurate detection of the target object by the target image processing model. Therefore, by training the initial image processing model using object region information, predicted object region information, detection results, sample labels, flipped detection results, and flipped sample labels, a target image processing model capable of more accurate object detection is obtained. Then, based on this target image processing model, object detection is performed on the image to be detected sent by the client, thereby obtaining accurate detection results, and the detection results are returned to the client for display.

[0339] The above is an illustrative scheme of another image processing apparatus according to this embodiment. It should be noted that the technical solution of the other image processing apparatus and the technical solution of the other image processing method described above belong to the same concept. For details not described in detail in the technical solution of the other image processing apparatus, please refer to the description of the technical solution of the other image processing method described above.

[0340] Corresponding to the above method embodiments, this specification also provides an embodiment of an information processing apparatus. Figure 13 A schematic diagram of the structure of an information processing apparatus according to one embodiment of this specification is shown. Figure 13 As shown, the device includes:

[0341] The image determination module 1302 is configured to determine a CT image to be detected of a target tumor located in the nasopharynx.

[0342] The object detection module 1304 is configured to perform tumor detection on the CT image to be detected using a target image processing model, and obtain the tumor detection result of the CT image to be detected.

[0343] The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result.

[0344] The information processing module 1306 is configured to determine the information processing result based on the tumor detection result of the CT image to be detected.

[0345] This specification provides an information processing apparatus that, during information processing, utilizes a large amount of target training data and enhanced training data to train an initial image processing model, thereby improving the accuracy of tumor detection results output by the target CT image processing model. Furthermore, during model training, to further improve tumor detection accuracy, the initial image processing model can be used to perform tumor detection on sample CT images and flipped sample CT images respectively, obtaining tumor detection results for the sample CT images and flipped tumor detection results for the flipped sample CT images. Based on the tumor detection results and flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined. Then, the tumor region information and predicted tumor region information are used to train the initial image processing model, thereby achieving accurate detection of the target tumor by the target image processing model. Therefore, by training the initial image processing model using tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor detection results, and flipped sample labels, a target image processing model capable of more accurate tumor detection is obtained. Subsequently, the accurate tumor detection results output by the target image processing model are processed to obtain accurate information processing results.

[0346] The above is an illustrative scheme of an information processing device according to this embodiment. It should be noted that the technical solution of this information processing device and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the information processing device, please refer to the description of the technical solution of the information processing method described above.

[0347] Figure 14 A structural block diagram of a computing device 1400 according to one embodiment of this specification is shown. The components of the computing device 1400 include, but are not limited to, a memory 1410 and a processor 1420. The processor 1420 is connected to the memory 1410 via a bus 1430, and a database 1450 is used to store data.

[0348] The computing device 1400 also includes an access device 1440, which enables the computing device 1400 to communicate via one or more networks 1460. Examples of such networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 1440 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0349] In one embodiment of this specification, the above-described components of the computing device 1400 and Figure 14 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 14 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0350] The computing device 1400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 1400 can also be a mobile or stationary server.

[0351] The processor 1420 is configured to execute a computer program / instruction that, when executed by the processor, implements the steps of any of the methods described above.

[0352] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the computing device embodiments are basically similar to any of the above-described method embodiments, so the description is relatively simple; relevant parts can be referred to in the description of any of the above-described method embodiments.

[0353] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0354] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the computer-readable storage medium embodiments are basically similar to any of the above-described method embodiments, so the description is relatively simple; relevant parts can be referred to in the description of any of the above-described method embodiments.

[0355] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0356] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solution of any of the above methods, and any details not described in detail in the technical solution of the computer program product can be referred to the description of the technical solution of any of the above methods.

[0357] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0358] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0359] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0360] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0361] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for training an image processing model, comprising: Determine the target training data and the augmented training data, wherein the target training data includes a sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object; and the augmented training data includes a flipped sample image corresponding to the sample image and a flipped sample label of the flipped sample image. Using the initial image processing model, object detection is performed on the sample image and the flipped sample image respectively to obtain the detection results of the sample image and the flipped detection results of the flipped sample image; Based on the detection results and the flip detection results, the predicted object region information of the target object in the sample image is determined, including: using a nonlinear processing unit to perform data augmentation on the detection results and the flip detection results to obtain augmented detection results and augmented flip detection results; comparing the object features corresponding to the target object contained in the augmented detection results and the flipped object features corresponding to the target object contained in the augmented flip detection results to determine the augmented image features corresponding to the sample image; and determining the predicted object region information of the target object in the sample image from the augmented image features. Based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label, the initial image processing model is trained to obtain the trained target image processing model.

2. The image processing model training method according to claim 1, wherein training the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flip sample label to obtain the trained target image processing model includes: Based on the object region information and the predicted object region information, a first loss function is calculated, wherein the first loss function is used to evaluate the difference between the object region information and the predicted object region information. Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a second loss function is calculated, wherein the second loss function is used to evaluate the difference between the detection results and the sample labels, and the difference between the flip detection results and the flipped sample labels; Based on the detection results, the sample labels, the flip detection results, and the flipped sample labels, a third loss function is calculated, wherein the third loss function is used to evaluate the similarity between the detection results and the sample labels, and the similarity between the flip detection results and the flipped sample labels; Based on the first loss function, the second loss function, and the third loss function, the model parameters of the initial image processing model are adjusted to obtain the trained target image processing model.

3. The image processing model training method according to claim 1, wherein comparing the object features corresponding to the target object contained in the enhancement detection result and the flipped object features corresponding to the target object contained in the enhancement flip detection result to determine the enhancement image features corresponding to the sample image includes: Using a nonlinear processing unit, it is determined that the object features contain multiple object feature points, and the flipped object features contain multiple flipped object feature points; From the plurality of flipped object feature points, determine the corresponding feature points of each object feature point, and determine the similarity between each object feature point and the corresponding feature point; If the similarity is greater than or equal to a first threshold, the similarity weight between the object feature point and the corresponding feature point is increased; if the similarity is less than the first threshold, the similarity weight between the object feature point and the corresponding feature point is decreased, thereby obtaining the enhanced image features corresponding to the sample image.

4. The image processing model training method according to claim 1, wherein determining the target training data and the augmented training data includes: Determine the target training data; The sample image is horizontally flipped to obtain the flipped sample image, and the sample label is horizontally flipped to obtain the flipped sample label of the flipped sample image; The flipped sample image and the flipped sample label are used as the augmented training data.

5. The image processing model training method according to claim 1, wherein the step of using an initial image processing model to perform object detection on the sample image and the flipped sample image respectively, to obtain the detection result of the sample image and the flipped detection result of the flipped sample image, includes: Using the encoder in the initial image processing model, the sample image and the flipped sample image are encoded respectively to obtain the sample image encoding of the sample image and the flipped sample image encoding of the flipped sample image; Using the decoder in the initial image processing model, the encoded sample image and the encoded flipped sample image are decoded respectively to obtain the detection result of the sample image and the flipped detection result of the flipped sample image.

6. The image processing model training method according to claim 1 or 4, wherein determining the target training data includes: Determine the object image of the target object located at the target location, and determine the object image to be processed from the object image, wherein the object image to be processed is any image in the object image; Based on the target region, a central axis is added to the image of the object to be processed to obtain an initial object image, wherein the central axis symmetrically divides the target region; Based on the initial object image, the central axis is added to other object images in the object image except for the object image to be processed, to obtain the target object image; The initial object image and the target object image are used as the sample images, and the sample labels corresponding to the sample images and the object region information of the target object are determined. The sample image, the sample label, and the object region information are used as the target training data.

7. The image processing model training method according to any one of claims 1 to 5, wherein the target location is the nasopharyngeal region, the target object is the target tumor, the sample image is a sample CT image, and the flipped sample image is a flipped sample CT image; The step of training the initial image processing model based on the object region information, the predicted object region information, the detection result, the sample label, the flip detection result, and the flipped sample label to obtain the trained target image processing model includes: Based on tumor region information, predicted tumor region information, tumor detection results, sample labels, flipped tumor results, and flipped sample labels, the initial image processing model is trained to obtain the trained target image processing model.

8. A method for training an image processing model, comprising: The target training data and the augmentation training data are determined. The target training data includes a sample CT image of a target tumor located in the nasopharynx, a sample label of the sample CT image, and tumor region information of the target tumor. The augmentation training data includes a flipped sample CT image corresponding to the sample CT image and a flipped sample label of the flipped sample CT image. Using an initial image processing model, tumor detection is performed on the sample CT image and the flipped sample CT image respectively to obtain the tumor detection results of the sample CT image and the flipped tumor detection results of the flipped sample CT image. Based on the tumor detection results and the flipped tumor detection results, the predicted tumor region information of the target tumor in the sample CT image is determined, including: using a nonlinear processing unit to perform data augmentation on the tumor detection results and the flipped tumor detection results to obtain augmented detection results and augmented flipped detection results; comparing the object features corresponding to the target tumor contained in the augmented detection results and the flipped object features corresponding to the target tumor contained in the augmented flipped detection results to determine the augmented image features corresponding to the sample CT image; and determining the predicted object region information of the target tumor in the sample CT image from the augmented image features. Based on the tumor region information, the predicted tumor region information, the tumor detection result, the sample label, the flipped tumor detection result, and the flipped sample label, the initial image processing model is trained to obtain the trained target image processing model.

9. An image processing method, comprising: Determine the image to be detected of the target object located at the target location; Object detection is performed on the image to be detected using a target image processing model to obtain the detection result of the image to be detected. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection results, sample labels, flip detection results, and flip sample labels. The target training data includes the sample image of the target object located at the target location, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by performing object detection on the sample image using the initial image processing model, and the flip detection result is determined by performing object detection on the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result. The predicted object region information is determined based on the detection results and the flip detection results, including: using a nonlinear processing unit to perform data augmentation on the detection results and the flip detection results to obtain augmented detection results and augmented flip detection results; comparing the object features corresponding to the target object contained in the augmented detection results and the flipped object features corresponding to the target object contained in the augmented flip detection results to determine the augmented image features corresponding to the sample image; and determining the predicted object region information of the target object in the sample image from the augmented image features.

10. An image processing method applied to a client of a medical system, comprising: In response to the user's selection operation on the client's display interface, the medical image to be detected is determined; The medical image to be detected is sent to the server of the medical system, and the detection result returned by the server is received. The detection result is the output result after object detection of the medical image to be detected by the target image processing model. The target image processing model is obtained by training an initial image processing model based on object region information determined by target training data and augmented training data, predicted object region information of the target object in the sample image, detection result, sample label, flip detection result, and flip sample label. The target training data includes the sample image of the target object located at the target site, the sample label of the sample image, and the object region information of the target object. The augmented training data includes the flipped sample image corresponding to the sample image and the flipped sample label of the flipped sample image. The detection result is determined by object detection of the sample image using the initial image processing model. The flip detection result is determined by object detection of the flipped sample image using the initial image processing model. The predicted object region information is determined based on the detection result and the flip detection result. The predicted object region information is determined based on the detection results and the flip detection results, including: using a nonlinear processing unit to perform data augmentation on the detection results and the flip detection results to obtain augmented detection results and augmented flip detection results; comparing the object features corresponding to the target object contained in the augmented detection results and the flipped object features corresponding to the target object contained in the augmented flip detection results to determine the augmented image features corresponding to the sample image; and determining the predicted object region information of the target object in the sample image from the augmented image features. The detection results are displayed to the user through the display interface.

11. A computer-aided diagnosis method for cancer, comprising: Identify the CT image of the target tumor located at the target site; A target image processing model is used to perform tumor detection on the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located at the target site, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by performing tumor detection on the sample CT image using the initial image processing model. The flipped tumor detection result is determined by performing tumor detection on the flipped sample CT image using the initial image processing model. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result. The predicted tumor region information is determined based on the tumor detection results and the flipped tumor detection results, including: using a nonlinear processing unit to perform data augmentation on the tumor detection results and the flipped tumor detection results to obtain augmented detection results and augmented flipped detection results; comparing the object features corresponding to the target tumor contained in the augmented detection results and the flipped object features corresponding to the target tumor contained in the augmented flipped detection results to determine the augmented image features corresponding to the sample CT image; and determining the predicted object region information of the target tumor in the sample CT image from the augmented image features.

12. A computer-aided diagnostic method for nasopharyngeal carcinoma, comprising: Identify the CT image of the target tumor located in the nasopharynx; A target image processing model is used to perform tumor detection on the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the nasopharynx, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by performing tumor detection on the sample CT image using the initial image processing model. The flipped tumor detection result is determined by performing tumor detection on the flipped sample CT image using the initial image processing model. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result. The predicted tumor region information is determined based on the tumor detection results and the flipped tumor detection results, including: using a nonlinear processing unit to perform data augmentation on the tumor detection results and the flipped tumor detection results to obtain augmented detection results and augmented flipped detection results; comparing the object features corresponding to the target tumor contained in the augmented detection results and the flipped object features corresponding to the target tumor contained in the augmented flipped detection results to determine the augmented image features corresponding to the sample CT image; and determining the predicted object region information of the target tumor in the sample CT image from the augmented image features.

13. A computer-aided diagnostic method for breast cancer, comprising: Identify the CT image of the target tumor located in the breast region; A target image processing model is used to detect tumors in the CT image to be detected, thereby obtaining the tumor detection result of the CT image to be detected. The target image processing model is obtained by training an initial image processing model based on tumor region information determined by target training data and enhanced training data, predicted tumor region information of the target tumor in the sample CT image, tumor detection result, sample label, flipped tumor detection result, and flipped sample label. The target training data includes the sample CT image of the target tumor located in the breast region, the sample label of the sample CT image, and the tumor region information of the target tumor. The enhanced training data includes the flipped sample CT image corresponding to the sample CT image and the flipped sample label of the flipped sample CT image. The tumor detection result is determined by using the initial image processing model to detect tumors in the sample CT image. The flipped tumor detection result is determined by using the initial image processing model to detect tumors in the flipped sample CT image. The predicted tumor region information is determined based on the tumor detection result and the flipped tumor detection result. The predicted tumor region information is determined based on the tumor detection results and the flipped tumor detection results, including: using a nonlinear processing unit to perform data augmentation on the tumor detection results and the flipped tumor detection results to obtain augmented detection results and augmented flipped detection results; comparing the object features corresponding to the target tumor contained in the augmented detection results and the flipped object features corresponding to the target tumor contained in the augmented flipped detection results to determine the augmented image features corresponding to the sample CT image; and determining the predicted object region information of the target tumor in the sample CT image from the augmented image features.

14. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 13.

15. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 13.

16. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 13.

Images