Intelligent head, neck, and pelvic cavity tumor image processing system

Abstract

This approach preserves minute tumor lesions while suppressing noise interference, designs gradient adjustment terms, reduces boundary offsets, improves segmentation reliability, avoids misclassification due to morphological differences, optimizes the detection of concurrent tumor scenes, and enhances tumor image recognition effectiveness. [Solution] In an intelligent head, neck, and pelvic cavity tumor image processing system, the image acquisition module collects historical head, neck, and pelvic cavity tumor images, tumor region masks, and tumor category labels to construct a tumor image set. The tumor image segmentation module constructs a tumor image segmentation model based on a flexible texture accumulation layer and a segmentation loss function for the tumor image set and the masks to obtain a tumor image identification image set. The tumor image identification module completes the construction of a tumor image identification model by performing secondary training after initial training using a primary loss function based on the identification image set. The tumor image processing module collects tumor CT scans in real time and sequentially inputs them into the segmentation model and the identification model to obtain processing results.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and more specifically to an intelligent head and neck and pelvic cavity tumor image processing system.

Background Art

[0002] A head and neck and pelvic cavity tumor image processing system is a specialized technical system that performs preprocessing, tumor region segmentation, feature analysis, and result visualization on CT medical images of head and neck and pelvic cavity tumors, providing data processing support. However, general tumor image processing systems have problems such as being likely to lose important details of micro tumor lesions, having blurred boundaries in the segmentation results and being prone to deviation, and having low reliability in segmentation. General tumor image processing systems also have problems such as being difficult to process different stages of the same tumor and the same imaging findings of different tumors, having confusion in judgment due to the overlap of lesion features, and further reducing the discrimination effect of tumor images.

Summary of the Invention

[0003] In view of the above situation, in order to overcome the drawbacks of the prior art, the present invention provides an intelligent head and neck and pelvic cavity tumor image processing system. For the problems in general tumor image processing systems, such as being likely to lose important details of micro tumor lesions, having blurred boundaries in the segmentation results and being prone to deviation, and having low reliability in segmentation, the present solution divides pixels according to tissue density, combines a flexible integration process, suppresses noise interference while retaining micro tumor lesions through the integration of local neighborhood information, designs a gradient adjustment term, reduces boundary offset, and further improves segmentation reliability. For the problems in general tumor image processing systems, such as being difficult to process different stages of the same tumor and the same imaging findings of different tumors, having confusion in judgment due to the overlap of lesion features, and further reducing the discrimination effect of tumor images, the present solution uses a feature generation evaluation unit and cross-modal fusion to avoid misjudgment due to morphological differences. Through secondary training, the learning of the model for rare tumors is strengthened, the determination of multiple tumor co-occurrence scenes is optimized based on candidate re-ranking and mixed loss, and the discrimination effect of tumor images is further enhanced.

[0004] The technical means employed in the present invention are as follows: The present invention is an intelligent head and neck / pelvic cavity tumor image processing system comprising an image acquisition module, a tumor image segmentation module, a tumor image identification module, and a tumor image processing module, The image acquisition module collects historical head and neck / pelvic tumor images, tumor area masks, and tumor category labels, and constructs a tumor image set. The tumor image segmentation module constructs a tumor image segmentation model based on a flexible texture integration layer and a segmentation loss function for a tumor image set and a tumor region mask, and obtains a tumor image identification image set. The tumor image recognition module, based on the tumor image recognition image set, undergoes preliminary training using a loss function, followed by secondary training to complete the construction of a tumor image recognition model. The tumor image processing module provides an intelligent head, neck, and pelvic cavity tumor image processing system that collects tumor CT images in real time, sequentially inputs them into a tumor image segmentation model and a tumor image recognition model, and obtains processing results.

[0005] Furthermore, in the image acquisition module, the chronological head and neck / pelvic tumor images include CT images of head and neck cancer and pelvic tumors, which are preprocessed to add information including tumor region masks and tumor category labels to the images, thereby obtaining a tumor image set.

[0006] Furthermore, the tumor image segmentation module includes network structure design and segmentation loss function design, specifically, In the aforementioned network structure design, the U-Net model of encoder-decoder architecture is adopted, and a flexible texture integration layer is introduced. In the aforementioned flexible texture integrating layer, a tumor region assignment function is defined, and tumor pixel values ​​are divided into three region categories. In the computation flow of the flexible texture integrating layer, local image blocks are extracted from the feature map of head, neck, and pelvic cavity tumor CT images, three sets of tumor region assignments are calculated for each pixel, and after joining, the summation score is obtained. The tumor region assignment set with the highest score is selected as the dominant one, and a dominant region feature block is generated. Finally, the flexible texture integrating layer outputs the processed head, neck, and pelvic cavity tumor CT image downsampled feature map. In the aforementioned partitioned loss function design, all pixels in a category are sorted in descending order of predicted probability, a cumulative error count and loss component are defined, the loss component is weighted, and the partitioned loss is obtained. After construction is complete, the tumor image set is processed to obtain a tumor image recognition image set.

[0007] Furthermore, the tumor image recognition module includes a model architecture component, the model including a feature extraction layer, a cross-modal fusion layer, a tumor feature generator, and a tumor category classifier. The aforementioned model architecture component is a functional unit, specifically, A feature generation and evaluation unit whose input is a combination of a CT image I and a tumor clinical template t(c), and whose output is the generation of a candidate sequence s. The category determination optimization unit extracts visual features from CT images I, obtains objective feature encodings, calculates the similarity between the objective feature encoding and the text description based on the text of the tumor feature description for each tumor category, selects candidate categories, and predicts the optimal label for the selected candidate categories using the feature generation and evaluation unit. The aforementioned preliminary discrimination loss includes a global loss and a local loss, the global loss uses forward cross-entropy, and the local loss uses binary cross-entropy. After the tumor image recognition set converges with training based on the preliminary discrimination loss function, the primary training is completed, and the tumor image recognition set is divided into a general tumor set and a rare tumor set, and secondary training is performed. This is a preliminary discrimination loss function design unit. For each image in a typical tumor set, a general tumor image recognition unit is input into a tumor image recognition model that has completed primary training, outputs a predicted triple, and defines accurate discrimination criteria. A rare tumor image recognition unit extracts all samples from the aforementioned rare tumor dataset as a test set, inputs each test image into a tumor image recognition model that has completed primary training, outputs a predicted triple, and defines accurate discrimination criteria. A common tumor set and a rare tumor set are selected, non-target tumor interference samples are added to construct a mixed interference dataset, the model generates five candidate triples for each CT image, the feature generation evaluation unit calculates the degree of image-text matching for each candidate label using the scoring criteria, the candidate label with the highest score and the candidate label with the lowest score are selected, and a candidate re-ranking evaluation unit calculates the re-ranking loss, This is used to realize a determination unit in which, when the tumor image recognition model converges to the re-ranking loss, secondary training is completed, and in this case the construction of the tumor image recognition model is completed.

[0008] Furthermore, the tumor image processing module collects images of head, neck, and pelvic cavity tumors in real time, sequentially inputs them into a tumor image segmentation model and a tumor image recognition model, and uses the image categories output by the tumor image recognition model as the tumor image processing results.

[0009] The effects obtained by using the above method are as follows: (1) In general tumor image processing systems, important details of minute tumor lesions are easily lost, the boundaries of the segmented results tend to be blurred and misaligned, and the reliability of segmentation is low. In response to these problems, this solution divides pixels by tissue density, combines a flexible accumulation process, preserves minute tumor lesions while suppressing noise interference through the accumulation of local neighborhood information, designs a gradient adjustment term to reduce boundary offset, and further improves segmentation reliability. (2) In general tumor image processing systems, it is difficult to process different stages of the same tumor and the same image findings of different tumors, and the overlap of lesion features leads to confusion in judgment and further reduces the tumor image recognition effect. This solution avoids misclassification due to morphological differences by using a feature generation and evaluation unit and cross-modal fusion. Secondary training enhances the model's learning of rare tumors, and the judgment of multi-tumor coexisting scenes is optimized based on candidate re-ranking and mixed loss, further improving the tumor image recognition effect. [Brief explanation of the drawing]

[0010] [Figure 1] This is a schematic flowchart of the intelligent head, neck, and pelvic cavity tumor image processing system provided by the present invention. [Figure 2] This is a schematic diagram of the functional units of the tumor image recognition module.

[0011] The accompanying drawings are provided to provide a further understanding of the present invention and constitute part of the specification, and are used to illustrate the present invention together with the embodiments of the present invention, and are not intended to limit the present invention. [Modes for carrying out the invention]

[0012] The technical means in embodiments of the present invention will be described clearly and completely below with reference to the drawings of the embodiments. Clearly, the embodiments described are not all embodiments of the present invention, but only a selection. All other embodiments that a person skilled in the art could obtain without creative work based on embodiments of the present invention are included within the scope of protection of the present invention.

[0013] In the description of this invention, terms indicating directions and positional relationships such as "up," "down," "front," "back," "left," "right," "top," "bottom," "inside," and "outside" are based on the directions and positional relationships shown in the drawings and are used to facilitate and simplify the explanation of this invention. They do not indicate or imply that a particular device or component has a particular direction or must be configured and operate in a particular direction. Therefore, they should not be understood as limiting the present invention.

[0014] [Example 1] Referring to Figure 1, the present invention is an intelligent head and neck / pelvic cavity tumor image processing system comprising an image acquisition module, a tumor image segmentation module, a tumor image identification module, and a tumor image processing module, The image acquisition module collects historical head, neck, and pelvic tumor images, tumor region masks, and tumor category labels, constructs a tumor image set, and transmits the data to the tumor image segmentation module. The tumor image segmentation module constructs a tumor image segmentation model based on a flexible texture integration layer and a segmentation loss function for the tumor image set and tumor region mask, obtains a tumor image identification image set, and transmits the data to the tumor image identification module. The tumor image recognition module, based on the tumor image recognition image set, performs a preliminary training using a loss function, then a secondary training to complete the construction of a tumor image recognition model, and transmits the data to the tumor image processing module. The tumor image processing module provides an intelligent head, neck, and pelvic cavity tumor image processing system that collects tumor CT images in real time, sequentially inputs them into a tumor image segmentation model and a tumor image recognition model, and obtains processing results.

[0015] [Example 2] Referring to FIG. 1, this embodiment is based on the above embodiment. The historical head and neck and pelvic cavity tumor images include CT images of head and neck cancers and pelvic cavity tumors, and preprocessing including format conversion (since the collected head and neck and pelvic cavity tumor images are in DICOM format, the format needs to be converted for subsequent segmentation operations), CT value calibration, and gray-scale normalization processing was performed. Marking information on the image includes a tumor region mask and a tumor category label included. The tumor region mask is binarized, where 1 indicates the tumor region and 0 indicates normal tissue. The tumor category label includes common tumors, rare tumors, and no tumors. Common tumors include nasopharyngeal carcinoma, laryngeal squamous cell carcinoma, hypopharyngeal carcinoma, cervical squamous cell carcinoma, rectal adenocarcinoma, and endometrial carcinoma. Rare tumors include olfactory neuroblastoma, mucoepidermoid carcinoma of the parotid gland, sacral chordoma, and vaginal clear cell carcinoma. The tumor features were described, and the description of the template referred to the "Diagnostic Criteria for Head and Neck Tumor Images (2022 Edition)" to obtain a tumor image set.

[0016] [Example 3] Referring to FIG. 1, this embodiment is based on the above embodiment. Based on the tumor image set and the tumor region mask, the tumor image segmentation module replaced the conventional max pooling in the downsampling of U-Net with a flexible texture integration layer to retain the details of the tumor boundary and micro-metastasis foci and constructed a tumor image segmentation model. The specific content is as follows. In the network structure design that employs the encoder-decoder architecture U-Net model, Encoder: Through 4 downsampling operations, the preprocessed head and neck and pelvic cavity tumor CT image was gradually reduced from 128×128×3 to 8×8×512. In each downsampling, after using a convolutional layer to extract features, a flexible texture integration layer was adopted instead of the conventional max pooling. Decoder architecture: Through 4 upsampling operations, the feature map was restored from 8×8×512 to 128×128×3, and finally a binary segmentation map was output. Skip connection: A direct connection of feature layers with the same resolution between the encoder and the decoder was constructed to directly transmit the details of the tumor margin extracted in the shallow layer of the encoder and the features of micro-lesions to the decoder side, and by supplementing the high-resolution shallow layer features, the blurring problem of the tumor boundary in the upsampling process was solved, and the spatial accuracy of tumor region segmentation was significantly improved. Introduction of flexible texture integration layer: Conventional max pooling is likely to lose important clinical details when processing head and neck and pelvic cavity tumor CT images. The flexible texture integration layer can retain these micro-features by integrating neighboring information and suppress irrelevant noises such as metal artifacts and gas artifacts. The specific content of the flexible texture integration layer is as follows. Definition of tumor region membership function: The tumor pixel values are divided into three region categories reflecting tissue characteristics, including normal tissue corresponding to low density and low signal, transition region corresponding to the tumor margin, and tumor core region corresponding to high density and high signal, and the gradient membership function is adopted.

Number

Number

Number

[0017] [Example 4] Referring to Figures 1 and 2, this embodiment is based on the above embodiment. The head and neck / pelvic tumor identification module constructs a head and neck / pelvic tumor image identification model by combining tumor category labels and tumor feature descriptions for a set of tumor image identification images. Specifically, it includes the following: Model Architecture Components: The model includes a feature extraction layer, a cross-modal fusion layer, a tumor feature generator, and a tumor category classifier. The feature extraction layer includes a visual encoder, a text encoder, and a feature alignment. The visual encoder employs ResNet-50 as its base backbone, removes the fully connected layer of the final layer, and outputs a 512-dimensional image feature vector containing tumor core visual information such as tumor morphology, CT-enhanced features, boundary features, and invasion extent. The text encoder encodes a tumor clinical template t(c) using BERT-base and outputs a 768-dimensional text feature vector containing semantic information of tumor pathology type, anatomical site, and typical image features. The feature alignment maps the visual features (512 dimensions) and text features (768 dimensions) to the same 640-dimensional feature space using a linear projection layer. The aforementioned cross-modal fusion layer employs the Transformer's cross-attention mechanism, using image features as keys and values ​​and text features as queries to achieve highly accurate association between visual details and text descriptions. The tumor feature generator, based on the GPT-2 architecture, generates structured tumor feature sequences s using cross-modal fusion features as input, and ensures consistency between the generated content and CT images and clinical templates through an autoregressive mechanism. The tumor category classifier performs fully connected layer mapping on cross-modal fusion features, outputs a category probability distribution for head and neck / pelvic tumors, and selects candidate categories by combining this with CLIP similarity scores, thereby improving the discrimination efficiency of scenes with multiple tumors occurring simultaneously. The aforementioned model architecture components are used to implement the following functional units. Feature generation and evaluation unit: Input: Combination of CT image I and tumor clinical template t(c); tumor clinical template is { <x>The head, neck, and pelvic cavity <y>It is located in the area, and the CT findings are <z>It is expressed as}, and among them, <x>This corresponds to the tumor pathology type, <y>These correspond to anatomical parts, <z>This corresponds to the core image features. Output: Candidate sequence s=<X:c,Y:p,Z:f> This generates the following: c corresponds to the predicted tumor pathology type, p to the predicted anatomical site, and f to the predicted core image features. Category Classification Optimization Unit: By optimizing the ranking of candidate labels, the model gives higher confidence to more reliable tumor identification results. Visual features are extracted from CT images I, convolutional features are extracted using ResNet-50, and an objective feature encoding including tumor morphology and enhancement information is obtained.

number

number

number

number

number

[0018] [Example 5] Referring to Figure 1, this embodiment is based on the above embodiment. The tumor image processing module collects images of head, neck, and pelvic cavity tumors, sequentially inputs them into a tumor image segmentation model and a tumor image recognition model, and the image categories output by the tumor image recognition model are used as the tumor image processing results.

[0019] Although embodiments of the present invention have been shown and described, it will be obvious to those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0020] Although the present invention and its embodiments have been described above, this description is not limiting, and the accompanying drawings represent only one embodiment of the present invention; the actual structure is not limited thereto. In other words, any similar modes and examples of the technical means designed by a person skilled in the art, without departing from the spirit of the invention and without creativity, should all fall within the scope of the present invention.< / z> < / y> < / x> < / z> < / y> < / x>

Claims

1. An intelligent head and neck / pelvic cavity tumor image processing system comprising an image acquisition module, a tumor image segmentation module, a tumor image identification module, and a tumor image processing module, The aforementioned image acquisition module collects historical head and neck / pelvic tumor images, tumor area masks, and tumor category labels, and constructs a tumor image set. The tumor image segmentation module constructs a tumor image segmentation model based on a flexible texture integration layer and a segmentation loss function for a tumor image set and a tumor region mask, and obtains a tumor image identification image set. The tumor image recognition module, based on the tumor image recognition image set, undergoes preliminary training using a loss function, followed by secondary training to complete the construction of a tumor image recognition model. The tumor image processing module is an intelligent head and neck / pelvic cavity tumor image processing system characterized by acquiring tumor CT images in real time, sequentially inputting them into a tumor image segmentation model and a tumor image recognition model, and obtaining processing results.

2. The aforementioned tumor image segmentation module includes network structure design and segmentation loss function design, and its specific details are as follows: In the aforementioned network structure design, the U-Net model of encoder-decoder architecture is adopted, and a flexible texture integration layer is introduced. In the flexible texture integrating layer, a tumor region assignment function is defined, and tumor pixel values ​​are divided into three region categories. In the computation flow of the flexible texture integrating layer, local image blocks are extracted from the feature map of head, neck, and pelvic cavity tumor CT images, the tumor region assignment of three sets for each pixel is calculated, and after joining, the sum is obtained to obtain an integrating score. The tumor region assignment set with the highest score is selected as the dominant one, and a dominant region feature block is generated. Finally, the flexible texture integrating layer outputs the processed CT image downsampled feature map of head, neck, and pelvic cavity tumors. The intelligent head and neck / pelvic cavity tumor image processing system according to claim 1, characterized in that the partition loss function design involves sorting all pixels of a category in descending order of predicted probability, defining a cumulative error count and loss component, weighting the loss component to obtain partition loss, processing the tumor image set after construction is complete, and obtaining a tumor image identification image set.

3. The tumor image recognition module includes a model architecture component, the model includes a feature extraction layer, a cross-modal fusion layer, a tumor feature generator, and a tumor category classifier. The aforementioned model architecture component consists of the following functional units, specifically: A feature generation and evaluation unit whose input is a combination of a CT image I and a tumor clinical template t(c), and whose output is the generation of a candidate sequence s, Category determination optimization unit, A basic discriminant loss function design unit, For each image in a typical tumor set, a general tumor image recognition unit is input into a tumor image recognition model that has completed primary training, outputs a predicted triple, and defines accurate discrimination criteria. A rare tumor image recognition unit extracts all samples from a rare tumor dataset as a test set, inputs each test image into a tumor image recognition model that has completed primary training, outputs a predicted triple, and defines accurate discrimination criteria. Candidate re-ranking evaluation unit, The intelligent head and neck / pelvic cavity tumor image processing system according to claim 2, characterized in that it is used to realize a determination unit in which secondary training is completed when the tumor image recognition model converges with respect to the reranking loss, in which case the construction of the tumor image recognition model is completed.

4. The intelligent head and neck / pelvic cavity tumor image processing system according to claim 3, characterized in that the category determination optimization unit extracts visual features from CT image I, obtains objective feature encoding, calculates the similarity between the objective feature encoding and the text description based on the tumor feature description text for each tumor category, selects candidate categories, and predicts the optimal label for the selected candidate categories using the feature generation evaluation unit.

5. The intelligent head and neck / pelvic cavity tumor image processing system according to claim 4, characterized in that, in the preliminary discrimination loss function design unit, the preliminary discrimination loss includes a global loss and a local loss, the global loss uses forward cross-entropy, the local loss uses binary cross-entropy, and after the tumor image discrimination image set converges through training based on the preliminary discrimination loss function, primary training is completed, the tumor image discrimination image set is divided into a general tumor set and a rare tumor set, and secondary training is performed.

6. The intelligent head and neck / pelvic cavity tumor image processing system according to claim 5, characterized in that the candidate re-ranking evaluation unit selects a common tumor set and a rare tumor set, adds non-target tumor interference samples to construct a mixed interference dataset, the model generates five candidate triples for each CT image, the feature generation evaluation unit calculates the image-text matching degree of each candidate label as a scoring criterion, selects the candidate label with the highest score and the candidate label with the lowest score, and calculates the re-ranking loss.

7. The intelligent head and neck / pelvic cavity tumor image processing system according to claim 6, characterized in that the tumor image processing module collects images of head and neck / pelvic cavity tumors, sequentially inputs them into a tumor image segmentation model and a tumor image recognition model, and uses the image categories output by the tumor image recognition model as the tumor image processing results.

Images