Multi-organ medical image segmentation method and device based on weak supervision

Abstract

The invention discloses a multi-organ medical image segmentation method and device based on weak supervision, relates to the technical field of image processing, and mainly aims to solve the problem that existing multi-organ medical image segmentation is poor in accuracy and effectiveness. Comprising the following steps: acquiring a collected multi-organ medical image, wherein the multi-organ medical image is obtained by computed tomography; performing standardized preprocessing on the multi-organ medical image; performing image segmentation on the multi-organ medical image subjected to standardization preprocessing based on a multi-organ segmentation network model subjected to model training to obtain a multi-organ image segmentation result; wherein the multi-organ segmentation network model is constructed by a double-branch network structure, the double-branch network structure comprises a shared encoder and two parallel decoding branches, and an overall loss function of the multi-organ segmentation network model comprises a cross entropy loss function configured with a dynamic threshold.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and apparatus for multi-organ medical image segmentation based on weak supervision. Background Technology

[0002] With the rapid development of artificial intelligence algorithms, deep learning algorithms have become the main image processing method for medical image recognition. In particular, in medical image segmentation, image processing methods such as thresholding, edge detection, and region growing can be used to adapt to different medical image segmentation tasks.

[0003] Currently, deep learning methods such as Convolutional Neural Networks (CNNs) and Transformers can achieve automatic feature extraction and rule-based segmentation. In organ segmentation of abdominal CT images, numerous CT image samples of various organ regions with pre-annotated graffiti are collected and trained using deep neural networks such as U-Net to complete the image segmentation task. For example, a dual-branch network (DB-Net) with a shared encoder and dual decoders can achieve mutual supervision by dynamically mixing the predictions of the two branches during training to generate soft pseudo-labels. However, this mutual supervision method still relies on the quality of branch prediction discrepancies and uncertainty estimation, and is too time-consuming in multi-organ scenarios on abdominal CT, thus affecting the accuracy and effectiveness of multi-organ medical image segmentation. Summary of the Invention

[0004] In view of this, this application provides a multi-organ medical image segmentation method and apparatus based on weak supervision, the main purpose of which is to solve the problems of poor accuracy and effectiveness of existing multi-organ medical image segmentation.

[0005] According to one aspect of this application, a weakly supervised multi-organ medical image segmentation method is provided, comprising: Acquire multi-organ medical images, which are obtained by computed tomography scans; The multi-organ medical images were subjected to standardized preprocessing. Based on the multi-organ segmentation network model that has completed model training, the multi-organ medical image that has completed standardized preprocessing is segmented to obtain the multi-organ image segmentation result. The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

[0006] Furthermore, before the multi-organ medical image that has undergone standardized preprocessing is segmented based on the multi-organ segmentation network model that has completed model training to obtain the multi-organ image segmentation result, the method further includes: Obtain multi-organ image training samples that include both graffiti annotations and un-graffiti annotations, wherein the multi-organ training samples are standardized image samples; Construct a dual-branch network structure that includes a shared encoder and two parallel decoding branches, wherein the decoding branches include a main branch decoder and an auxiliary branch decoder; The dual-branch network structure is trained using the multi-organ graph training samples, and the model training is completed when the loss value calculated by the overall loss function meets the preset loss threshold, thus obtaining the multi-organ segmentation network model.

[0007] Furthermore, the method also includes: The prediction accuracy is obtained by performing a nonlinear transformation based on the organ category and initial threshold of multiple organs. A category accuracy expression is constructed based on the number of pseudo-label pixels corresponding to the organ type and the number of background categories. The number of pseudo-label pixels is then normalized and solved. The expression of the number of pseudo-label pixels is constructed based on the main branch prediction map, indicator function, and dynamic threshold. A cross-entropy loss function is constructed based on the pseudo-label loss, the prediction accuracy, and the calculated pseudo-label pixel count.

[0008] Further, the step of training the dual-branch network structure using the multi-organ map training samples, and completing the model training when the loss value calculated by the overall loss function meets a preset loss threshold, to obtain the multi-organ segmentation network model, includes: Feature representations are obtained by extracting features from graffiti annotation samples in the multi-organ image training samples using the shared encoder; The shared features of the feature representation are decoded by the main branch decoder and the auxiliary branch decoder to obtain a first probability map and a second probability map; The pseudo-label pixels are determined by weighted mixing of the first probability map and the second probability map using a random mixing method. The pseudo-label loss is determined based on the pseudo-label pixels and the main branch prediction map.

[0009] Furthermore, the overall loss function includes a decoder loss function, a loss term, and a cross-entropy loss function configured with a dynamic threshold.

[0010] Furthermore, the method also includes: Acquire historical 3D medical images; The organ parts in the historical 3D images are marked with graffiti to obtain graffiti-annotated samples, and the unmarked samples are filtered out. The graffiti-annotated samples and the un-graffiti-annotated samples are standardized and preprocessed to generate the multi-organ image training samples.

[0011] Furthermore, the standardization preprocessing of the multi-organ medical images includes: The multi-organ medical images are standardized according to a preset voxel unit range, and the standardized multi-organ medical images are then normalized. The normalized multi-organ medical images are cropped according to the sample size to obtain preprocessed multi-organ medical images.

[0012] According to another aspect of this application, a multi-organ medical image segmentation device based on weak supervision is provided, comprising: The acquisition module is used to acquire collected multi-organ medical images, which are obtained by computed tomography scans. The processing module is used to perform standardized preprocessing on the multi-organ medical images; The segmentation module is used to perform image segmentation on the multi-organ medical image that has undergone standardized preprocessing based on the multi-organ segmentation network model that has been trained, so as to obtain the multi-organ image segmentation result. The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

[0013] Furthermore, the device also includes: a construction module and a training module. The acquisition module is further configured to acquire training samples of multi-organ images containing graffiti annotations and un-graffiti annotations, wherein the multi-organ training samples are standardized image samples; The building module is used to build a dual-branch network structure that includes a shared encoder and two parallel decoding branches, wherein the decoding branches include a main branch decoder and an auxiliary branch decoder. The training module is used to train the dual-branch network structure using the multi-organ graph training samples, and completes the model training when the loss value calculated by the overall loss function meets the preset loss threshold, thereby obtaining a multi-organ segmentation network model.

[0014] Furthermore, the device also includes: The construction module is further configured to perform nonlinear transformation based on organ categories and initial thresholds of multiple organs to obtain prediction accuracy; construct a category accuracy expression based on the pseudo-label pixel count expression corresponding to the organ type and the background category count, and normalize the pseudo-label pixel count expression, which is constructed based on the main branch prediction map, indicator function, and dynamic threshold; and construct a cross-entropy loss function based on the pseudo-label loss, the prediction accuracy, and the obtained pseudo-label pixel count expression.

[0015] Furthermore, The training module is further configured to extract features from the graffiti annotation samples in the multi-organ image training samples using the shared encoder to obtain feature representations; decode the shared features of the feature representations using the main branch decoder and the auxiliary branch decoder to obtain a first probability map and a second probability map; perform weighted mixing of the first probability map and the second probability map using a random mixing method to determine pseudo-label pixels; and determine the pseudo-label loss based on the pseudo-label pixels and the main branch prediction map.

[0016] Furthermore, the overall loss function includes a decoder loss function, a loss term, and a cross-entropy loss function configured with a dynamic threshold.

[0017] Furthermore, The acquisition module is also used to acquire historical three-dimensional medical images; to mark the organ parts in the historical three-dimensional images in a graffiti style to obtain graffiti-annotated samples, and to filter out unmarked samples; to perform standardized preprocessing on the graffiti-annotated samples and the unmarked samples to generate the multi-organ image training samples.

[0018] Furthermore, The processing module is further configured to standardize the multi-organ medical image according to a preset voxel unit range, normalize the standardized multi-organ medical image, and crop the normalized multi-organ medical image according to the sample size to obtain a preprocessed multi-organ medical image.

[0019] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the above-described weakly supervised multi-organ medical image segmentation method.

[0020] According to another aspect of this application, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described weakly supervised multi-organ medical image segmentation method.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a weakly supervised multi-organ medical image segmentation method and apparatus. Compared with the prior art, the embodiments of this application acquire multi-organ medical images obtained by computed tomography scans; perform standardized preprocessing on the multi-organ medical images; and perform image segmentation on the standardized preprocessed multi-organ medical images based on a multi-organ segmentation network model that has completed model training to obtain multi-organ image segmentation results. The multi-organ segmentation network model is constructed with a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. Furthermore, the overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold, achieving a dynamic balance between quality and quantity during the pseudo-label generation stage. This avoids the defects of the fixed threshold mechanism, significantly reduces the interference of low-confidence pseudo-labels on gradient propagation, and thus improves image segmentation accuracy.

[0022] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of a multi-organ medical image segmentation method based on weak supervision, according to an embodiment of this application. Figure 2 A schematic diagram of a dual-branch network structure provided in an embodiment of this application is shown; Figure 3 An embodiment of this application is shown. Diagram illustrating the construction of the loss function; Figure 4 This illustration shows a block diagram of a multi-organ medical image segmentation device based on weak supervision, as provided in an embodiment of this application. Figure 5A schematic diagram of the structure of a terminal provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] This application provides a weakly supervised multi-organ medical image segmentation method, such as... Figure 1 As shown, the method includes: 101. Acquire medical images of multiple organs.

[0027] In this embodiment of the application, the multi-organ medical images are obtained by computed tomography, i.e., multi-organ CT images. Preferably, the abdomen of the user is scanned by a CT device to obtain an abdominal CT image, so as to perform image segmentation and identify abdominal organs.

[0028] It should be noted that the current execution end, as the main body for image segmentation, can be a server, such as a cloud server, or a terminal device, to perform image segmentation after acquiring medical images of multiple organs.

[0029] 102. Perform standardized preprocessing on the multi-organ medical images.

[0030] In this embodiment of the application, after the current execution end obtains the multi-organ medical image to be segmented, in order to adapt to the model for segmentation, the multi-organ medical image is subjected to standardized preprocessing, which may include standardization and normalization of voxel range, etc., to adapt to the input requirements of the model network. This embodiment of the application does not make specific limitations.

[0031] 103. Based on the multi-organ segmentation network model that has completed model training, the multi-organ medical image that has completed standardized preprocessing is segmented to obtain the multi-organ image segmentation result.

[0032] In this embodiment of the application, the current execution end inputs the standardized preprocessed multi-organ medical image into the multi-organ segmentation network. At this time, the multi-organ segmentation network model is constructed with a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold, so that model training can be completed when the loss value calculated by the overall loss function reaches the requirement.

[0033] It should be noted that, in this embodiment, the dual-branch network structure (DB-Net) is a weakly supervised learning framework that includes a shared encoder and two parallel decoding branches. In this case, the decoding branches include a main branch decoder and an auxiliary branch decoder, which perform decoding separately, as shown below. Figure 2 As shown.

[0034] In another embodiment of this application, for further definition and explanation, before the step of performing image segmentation on the multi-organ medical image that has undergone standardized preprocessing based on the multi-organ segmentation network model that has completed model training, and obtaining the multi-organ image segmentation result, the method further includes: Obtain training samples of multi-organ images containing both graffiti annotations and un-graffiti annotations; Construct a two-branch network structure that includes a shared encoder and two parallel decoding branches; The dual-branch network structure is trained using the multi-organ graph training samples, and the model training is completed when the loss value calculated by the overall loss function meets the preset loss threshold, thus obtaining the multi-organ segmentation network model.

[0035] To increase the diversity and stability of the model's prediction results, thereby improving the accuracy and effectiveness of image segmentation, the current execution end pre-trains the constructed dual-branch network structure. The decoding branch includes a main branch decoder and an auxiliary branch decoder. The main branch encoder extracts the 3D feature representation of the input image samples, while the main branch decoder and auxiliary branch decoder generate segmentation probability maps and obtain pseudo-labels through a random mixing mechanism. The overall loss function includes a decoder loss function, a loss term, and a cross-entropy loss function configured with a dynamic threshold. Here, the loss term is used to balance the impact of the loss function on model training and can be preferentially configured to 8; however, this embodiment does not impose specific limitations.

[0036] In a specific implementation scenario, the total loss function is expressed as: ; in, Let the decoder loss function be... For cross-entropy loss functions configured with dynamic thresholds, The loss term is preferably configured to be 8, which enables the model to fully utilize both labeled and unlabeled data for learning and optimization. This application does not impose specific limitations on the embodiments. It should be noted that the multi-organ training samples are standardized image samples, which can be constructed by selecting and annotating historical 3D medical images. For example, annotating specific abdominal organ images can be used as training samples. This application does not impose any specific limitations on these samples.

[0037] In some embodiments, for constructing the decoder loss function, first define Let and represent all image samples and their corresponding graffiti annotation samples in the training set, respectively. The number of training images, , indicating the training set The complete set of image pixels in the image. and These represent the labeled and unlabeled areas, respectively. Simultaneously, set... This represents the number of organ categories segmented. At this point, the image's annotations... It can contain There are three categories, with an additional category representing unlabeled pixels. During training, each training step samples the labeled doodle data and feeds it into the network for processing. For labeled image pixels in the doodle, cross-entropy (pCE) loss can be applied to each decoder; that is, during gradient backpropagation, unlabeled pixels can be ignored, and only labeled pixels are considered to guide model training. In this case, the decoder loss function can be expressed as: ; in, and These are respectively represented as the pixels predicted by the two branches. Category The probability, It's graffiti markings The labeled pixel sample set in For pixel p, assign a one-hot encoding to the category s corresponding to the graffiti annotation (if pixel p is labeled as category s, then...). (If it is 1, otherwise it is 0).

[0038] In another embodiment of this application, for further definition and explanation, the steps also include: The prediction accuracy is obtained by performing a nonlinear transformation based on the organ category and initial threshold of multiple organs. A category accuracy expression is constructed based on the number of pseudo-label pixels corresponding to the organ type and the number of background categories, and the number of pseudo-label pixels is normalized and solved. A cross-entropy loss function is constructed based on the pseudo-label loss, the prediction accuracy, and the calculated pseudo-label pixel count.

[0039] To achieve adaptive setting of dynamic thresholds for different organ categories during training—for example, setting higher thresholds for large organs to ensure pseudo-label quality and lower thresholds for small organs to ensure sufficient supervision quantity—and thus achieving a dynamic balance of supervision signals, the current execution-side model, before training, sets the organ category 's' and initial thresholds for multiple organs. After performing a nonlinear transformation, the prediction accuracy is obtained. The formula for calculating the prediction accuracy is then expressed as: ; in, At time t, the initial threshold The preferred configuration is 0.95. Indicates organ category At the iteration time The prediction accuracy This is a non-linear transformation function used for non-linear transformation. In this case, the threshold can be adjusted according to the learning state of different categories, making the selection of pseudo-labels for different categories more reasonable. Furthermore, the non-linear function... Represented as: , By using nonlinear transformation, the problem of large fluctuations in the threshold during the early stages of training can be avoided. Compared with linear functions, it has higher stability, thus enabling the dynamic threshold to gradually converge as the class learning progresses throughout the training process.

[0040] Furthermore, the current execution end can also construct a category accuracy expression based on the number of pseudo-label pixels corresponding to organ types and the number of background categories, thereby normalizing the number of pseudo-label pixels. At this point, the category accuracy expression... It can be represented as: ; in, In the iteration time At that time, category The predicted probability exceeds the initial threshold. The number of pixels in the pseudo-label. For the number of background categories, This represents the total number of categories. Furthermore, by normalizing the number of pseudo-labels for each category to the background class, we can avoid the model's predictions being heavily biased towards one category in the early stages of training, thus obtaining a more stable threshold estimate. Additionally, the number of pseudo-label pixels represents... Represented as: ; in, The size of a sample image batch. The main branch at the iteration time The predicted probability map This is an indicator function; at this point, when the predicted probability of a certain voxel is greater than the initial threshold... And its prediction category is At that time, the voxel was included The number of pixels is determined. Finally, a cross-entropy loss function is constructed based on the pseudo-label loss, prediction accuracy, and the obtained expression for the number of pseudo-label pixels. The expression for the number of pseudo-label pixels is constructed based on the main branch prediction map, indicator function, and dynamic threshold. Furthermore, a loss function is constructed based on the dynamic threshold. That is, the cross-entropy function can be expressed as: ; in, This can be represented as a dynamic threshold mask, such as Figure 3 As shown, the threshold masking mechanism increases the contribution of small organs to the training loss, thereby improving the overall segmentation model's ability to learn about small organs, taking into account the characteristic that small organs are more easily overlooked.

[0041] In another embodiment of this application, for further definition and explanation, the step of training the dual-branch network structure using the multi-organ map training samples, and completing the model training to obtain the multi-organ segmentation network model when the loss value calculated by the overall loss function meets a preset loss threshold, includes: Feature representations are obtained by extracting features from graffiti annotation samples in the multi-organ image training samples using the shared encoder; The shared features of the feature representation are decoded by the main branch decoder and the auxiliary branch decoder to obtain a first probability map and a second probability map; The pseudo-label pixels are determined by weighted mixing of the first probability map and the second probability map using a random mixing method. The pseudo-label loss is determined based on the pseudo-label pixels and the main branch prediction map.

[0042] To achieve the training objective of the weakly supervised model, and to improve image segmentation accuracy by utilizing pseudo-labels for unlabeled regions for quality screening and supervision, the current execution end extracts features from the graffiti annotation samples in the multi-organ image training samples using a shared encoder. This process involves extracting multi-scale feature representations through the shared encoder to obtain shared features. Simultaneously, the main branch decoder and the auxiliary branch decoder decode the shared features, outputting the corresponding first probability map. With the second probability diagram Furthermore, the first probability map and the second probability map are weighted and mixed using a random mixing method to determine the pseudo-label pixels. That is, by introducing a random mixing mechanism, the outputs of the main branch decoder and the auxiliary branch decoder are weighted according to random coefficients (such as...). Weighted fusion is performed to obtain pseudo-labels. This is a type of soft pseudo-label, where the formula for generating pseudo-label pixels is expressed as: At this point, by increasing diversity while maintaining consistency in prediction results, the construction effectively improves the stability and robustness of pseudo-labels, providing a foundation for subsequent dynamic threshold pseudo-label selection and loss calculation. Finally, the pseudo-label loss is determined based on the pseudo-label pixels and the main branch prediction overview, and can be expressed as: , This is a soft pseudo-label obtained by linearly mixing the probability maps output by the two branches (main branch and auxiliary branch) according to a random coefficient β pixel by pixel at the t-th iteration. Main branch decoder (the first probability graph corresponding to the output of the main branch decoder) The probability map output at the t-th iteration.

[0043] In another embodiment of this application, for further definition and explanation, the steps also include: Acquire historical 3D medical images; The organ parts in the historical 3D images are marked with graffiti to obtain graffiti-annotated samples, and the unmarked samples are filtered out. The graffiti-annotated samples and the un-graffiti-annotated samples are standardized and preprocessed to generate the multi-organ image training samples.

[0044] To reduce small organ segmentation errors by constructing training samples using scribble-style weak annotation, the current execution end first acquires historical 3D medical images. This can be obtained from CT scans of different body parts from medical enterprises or public medical image databases. Organs are then labeled using scribble-style annotation. This can be done using a Python script, where annotations cover only a small portion of pixels, leaving the rest unlabeled, thus forming training samples. Furthermore, to adapt to model training, the labeled samples are preprocessed to obtain multi-organ image training samples. Preprocessing methods can include standardization, normalization, and random cropping, etc., which are not specifically limited in this embodiment.

[0045] In another embodiment of this application, for further definition and explanation, the step of standardizing preprocessing the multi-organ medical images includes: The multi-organ medical images are standardized according to a preset voxel unit range, and the standardized multi-organ medical images are then normalized. The normalized multi-organ medical images are cropped according to the sample size to obtain preprocessed multi-organ medical images.

[0046] To enable effective model processing of multi-organ medical images and improve data processing efficiency, the current execution end pre-processes the acquired multi-organ medical images using standardization. Specifically, firstly, the multi-organ medical images are standardized according to a preset voxel unit range. Then, the standardized images are normalized, meaning the CT voxel values ​​of the multi-organ medical images are converted to a preset voxel unit range, preferably Hounsfield units of [–400, 400], to highlight organ tissues and suppress background interference. Next, the voxel values ​​of the standardized multi-organ medical images are normalized, such as to the [0, 1] interval, using the Min-Max standardization method; this embodiment does not specify a particular method. Finally, the normalized multi-organ medical images are cropped according to the sample size, such as randomly cropping them to a sample size of 80×96×96, to obtain the input for network model training, thus adapting to the network input size.

[0047] This application provides a weakly supervised multi-organ medical image segmentation method. Compared with the prior art, this application acquires multi-organ medical images obtained by computed tomography scans; performs standardized preprocessing on the multi-organ medical images; and segments the standardized preprocessed multi-organ medical images based on a multi-organ segmentation network model that has completed model training to obtain multi-organ image segmentation results. The multi-organ segmentation network model is constructed with a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. Furthermore, the overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold, achieving a dynamic balance between quality and quantity during the pseudo-label generation stage. This avoids the defects of the fixed threshold mechanism, significantly reduces the interference of low-confidence pseudo-labels on gradient propagation, and thus improves image segmentation accuracy.

[0048] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this application provides a multi-organ medical image segmentation device based on weak supervision, such as... Figure 4 As shown, the device includes: The acquisition module 21 is used to acquire acquired multi-organ medical images, which are obtained by computed tomography scans. Processing module 22 is used to perform standardized preprocessing on the multi-organ medical images; Segmentation module 23 is used to perform image segmentation on the multi-organ medical image that has completed standardized preprocessing based on the multi-organ segmentation network model that has completed model training, so as to obtain the multi-organ image segmentation result; The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

[0049] Furthermore, the device also includes: a construction module and a training module. The acquisition module is further configured to acquire training samples of multi-organ images containing graffiti annotations and un-graffiti annotations, wherein the multi-organ training samples are standardized image samples; The building module is used to build a dual-branch network structure that includes a shared encoder and two parallel decoding branches, wherein the decoding branches include a main branch decoder and an auxiliary branch decoder. The training module is used to train the dual-branch network structure using the multi-organ graph training samples, and completes the model training when the loss value calculated by the overall loss function meets the preset loss threshold, thereby obtaining a multi-organ segmentation network model.

[0050] Furthermore, the device also includes: The construction module is further configured to perform nonlinear transformation based on organ categories and initial thresholds of multiple organs to obtain prediction accuracy; construct a category accuracy expression based on the pseudo-label pixel count expression corresponding to the organ type and the background category count, and normalize the pseudo-label pixel count expression, which is constructed based on the main branch prediction map, indicator function, and dynamic threshold; and construct a cross-entropy loss function based on the pseudo-label loss, the prediction accuracy, and the obtained pseudo-label pixel count expression.

[0051] Furthermore, The training module is further configured to extract features from the graffiti annotation samples in the multi-organ image training samples using the shared encoder to obtain feature representations; decode the shared features of the feature representations using the main branch decoder and the auxiliary branch decoder to obtain a first probability map and a second probability map; perform weighted mixing of the first probability map and the second probability map using a random mixing method to determine pseudo-label pixels; and determine the pseudo-label loss based on the pseudo-label pixels and the main branch prediction map.

[0052] Furthermore, the overall loss function includes a decoder loss function, a loss term, and a cross-entropy loss function configured with a dynamic threshold.

[0053] Furthermore, The acquisition module is also used to acquire historical three-dimensional medical images; to mark the organ parts in the historical three-dimensional images in a graffiti style to obtain graffiti-annotated samples, and to filter out unmarked samples; to perform standardized preprocessing on the graffiti-annotated samples and the unmarked samples to generate the multi-organ image training samples.

[0054] Furthermore, The processing module is further configured to standardize the multi-organ medical image according to a preset voxel unit range, normalize the standardized multi-organ medical image, and crop the normalized multi-organ medical image according to the sample size to obtain a preprocessed multi-organ medical image.

[0055] This application provides a weakly supervised multi-organ medical image segmentation device. Compared with the prior art, this application acquires multi-organ medical images obtained by computed tomography scans; performs standardized preprocessing on the multi-organ medical images; and segments the standardized preprocessed multi-organ medical images based on a multi-organ segmentation network model that has completed model training to obtain multi-organ image segmentation results. The multi-organ segmentation network model is constructed with a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. Furthermore, the overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold, achieving a dynamic balance between quality and quantity during the pseudo-label generation stage. This avoids the defects of the fixed threshold mechanism, significantly reduces the interference of low-confidence pseudo-labels on gradient propagation, and thus improves image segmentation accuracy.

[0056] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the weakly supervised multi-organ medical image segmentation method in any of the above method embodiments.

[0057] Figure 5 The diagram shows a structural schematic of a terminal according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the terminal.

[0058] like Figure 5 As shown, the terminal may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0059] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0060] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0061] The processor 302 is used to execute program 310, which can specifically execute the relevant steps in the above embodiments of the multi-organ medical image segmentation method based on weak supervision.

[0062] Specifically, program 310 may include program code that includes computer operation instructions.

[0063] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The terminal includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0064] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0065] Specifically, program 310 can be used to cause processor 302 to perform the following operations: Acquire multi-organ medical images, which are obtained by computed tomography scans; The multi-organ medical images were subjected to standardized preprocessing. Based on the multi-organ segmentation network model that has completed model training, the multi-organ medical image that has completed standardized preprocessing is segmented to obtain the multi-organ image segmentation result. The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

[0066] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0067] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A multi-organ medical image segmentation method based on weak supervision, characterized in that, include: Acquire multi-organ medical images, which are obtained by computed tomography scans; The multi-organ medical images were subjected to standardized preprocessing. Based on the multi-organ segmentation network model that has completed model training, the multi-organ medical image that has completed standardized preprocessing is segmented to obtain the multi-organ image segmentation result. The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

2. The method according to claim 1, characterized in that, Before the multi-organ segmentation network model, which has already undergone model training, performs image segmentation on the standardized preprocessed multi-organ medical image to obtain the multi-organ image segmentation result, the method further includes: Obtain multi-organ image training samples that include graffiti annotations and un-graffiti annotations, wherein the multi-organ image training samples are standardized image samples; Construct a dual-branch network structure that includes a shared encoder and two parallel decoding branches, wherein the decoding branches include a main branch decoder and an auxiliary branch decoder; The dual-branch network structure is trained using the multi-organ graph training samples, and the model training is completed when the loss value calculated by the overall loss function meets the preset loss threshold, thus obtaining the multi-organ segmentation network model.

3. The method according to claim 2, characterized in that, The method further includes: The prediction accuracy is obtained by performing a nonlinear transformation based on the organ category and initial threshold of multiple organs. A category accuracy expression is constructed based on the number of pseudo-label pixels corresponding to the organ category and the number of background categories. The number of pseudo-label pixels is then normalized and solved. The expression of the number of pseudo-label pixels is constructed based on the main branch prediction map, indicator function, and dynamic threshold. A cross-entropy loss function is constructed based on the pseudo-label loss, the prediction accuracy, and the calculated pseudo-label pixel count.

4. The method according to claim 3, characterized in that, The step of training the dual-branch network structure using the multi-organ map training samples, and completing the model training when the loss value calculated by the overall loss function meets a preset loss threshold, to obtain the multi-organ segmentation network model includes: The shared encoder is used to extract features from the graffiti annotation samples in the multi-organ image training samples to obtain feature representations; The shared features of the feature representation are decoded by the main branch decoder and the auxiliary branch decoder to obtain a first probability map and a second probability map; The pseudo-label pixels are determined by weighted mixing of the first probability map and the second probability map using a random mixing method. The pseudo-label loss is determined based on the pseudo-label pixels and the main branch prediction map.

5. The method according to claim 2, characterized in that, The overall loss function includes the decoder loss function, the loss term, and the cross-entropy loss function configured with a dynamic threshold.

6. The method according to claim 2, characterized in that, The method further includes: Acquire historical 3D medical images; The organs in the historical 3D medical images were marked with graffiti to obtain graffiti-marked samples, and the unmarked samples were filtered out. The graffiti-annotated samples and the un-graffiti-annotated samples are standardized and preprocessed to generate the multi-organ image training samples.

7. The method according to claim 1, characterized in that, The standardization preprocessing of the multi-organ medical images includes: The multi-organ medical images are standardized according to a preset voxel unit range, and the standardized multi-organ medical images are then normalized. The normalized multi-organ medical images are cropped according to the sample size to obtain preprocessed multi-organ medical images.

8. A multi-organ medical image segmentation device based on weak supervision, characterized in that, include: The acquisition module is used to acquire collected multi-organ medical images, which are obtained by computed tomography scans. The processing module is used to perform standardized preprocessing on the multi-organ medical images; The segmentation module is used to perform image segmentation on the multi-organ medical image that has undergone standardized preprocessing based on the multi-organ segmentation network model that has been trained, so as to obtain the multi-organ image segmentation result. The multi-organ segmentation network model is constructed using a dual-branch network structure, which includes a shared encoder and two parallel decoding branches. The overall loss function of the multi-organ segmentation network model includes a cross-entropy loss function configured with a dynamic threshold.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images