Image segmentation model training method based on bidirectional hierarchical self-distillation and related device

Abstract

The invention discloses an image segmentation model training method based on bidirectional hierarchical self-distillation and a related device. The method comprises the following steps: constructing an image segmentation model; and training the image segmentation model by using the Dice loss function and the KL divergence bidirectional distillation loss to obtain a trained image segmentation model, and the method and the related device can improve the precision of the image segmentation model.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and medical image processing technology, and relates to an image segmentation model training method and related apparatus based on bidirectional hierarchical self-distillation. Background Technology

[0002] Existing knowledge distillation (KD) methods are widely used to improve the performance of lightweight models, especially in the field of semantic segmentation. However, they generally rely on: deeply trained and structurally complex teacher models; costly distillation data preparation; and multi-stage training processes (requiring the teacher model to be trained first, followed by the student model).

[0003] However, in real-world medical imaging scenarios, data annotation is scarce and models need to be deployed on edge devices, making traditional distillation methods difficult to apply.

[0004] Knowledge distillation is a model compression technique that uses teacher networks to guide student network learning. Mainstream methods include:

[0005] The Soft Targets distillation method proposed by Hinton et al. [Hinton, GE, Vinyals, O., & Dean, J. (2015). Distilling the Knowledge in a Neural Network. ArXiv, abs / 1503.02531.]

[0006] Attention Transfer[Guo,Z.,Yan,H.,Li,H.,&Lin,X.(2023).Class AttentionTransfer Based Knowledge Distillation.2023IEEE / CVF Conference on ComputerVision and Pattern Recognition(CVPR),11868-11877.], FitNet[Romero,A.,Ballas,N.,Kahou,SE,Chassang,A.,Gatta,C.,&Bengio,Y.(2014).FitNets:Hints for ThinDeep Nets.CoRR,abs / 1412.6550.] and other structural layer distillation methods;

[0007] CIRKD [Yang,C.,Zhou,H.,An,Z.,Jiang,X.,Xu,Y.,&Zhang,Q.(2022). Cross-Image Relational Knowledge Distillation for Semantic Segmentation.2022IEEE / CVF Conference on Computer Vision and Pattern Recognition(CVPR),12309-12318.], DecoupleKD [Zhao,B.,Cui,Q.,Song,R.,Qiu,Y.,&Liang,J.(2022).DecoupledKnowledge Distillation.2022IEEE / CVF Conference on Computer Vision and Pattern Recognition(CVPR),11943-11952.] and other distillation methods for segmentation tasks;

[0008] In recent years, self-distillation has gradually gained popularity, with representative methods including:

[0009] BYOT (own teacher) method [Zhang, L., Song, J., Gao, A., Chen, J., Bao, C., & Ma, K. (2019). Be Your Own Teacher: Improve the Performance of Convolutional NeuralNetworks via Self Distillation. 2019IEEE / CVF International Conference on ComputerVision (ICCV), 3712-3721.];

[0010] Self-Distillation method [Zhang, L., Bao, C., & Ma, K. (2021). Self-Distillation: Towards Efficient and Compact Neural Networks. IEEE Transactions onPatternAnalysis and Machine Intelligence, 44, 4388-4403.];

[0011] Distillation from the Last Mini-Batch (DLB) method [Shen, Y., Xu, L., Yang, Y., Li, Y., & Guo, Y. (2022). Self-Distillation from the Last Mini-Batch for Consistency Regularization. 2022 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11933-11942.];

[0012] However, most of these methods only perform one-way information transmission, that is, high-level prediction guides low-level prediction, ignoring the reverse influence of low-level structural information on the accuracy of high-level prediction, which ultimately leads to a decrease in the segmentation accuracy of the image segmentation model. Summary of the Invention

[0013] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and apparatus for training an image segmentation model based on bidirectional hierarchical self-distillation, which can improve the accuracy of the image segmentation model.

[0014] To achieve the above objectives, this invention discloses a method for training an image segmentation model based on bidirectional hierarchical self-distillation, comprising:

[0015] Construct an image segmentation model;

[0016] The image segmentation model is trained using the Dice loss function and the KL divergence bidirectional distillation loss to obtain the trained image segmentation model.

[0017] A further improvement of the image segmentation model training method based on bidirectional hierarchical self-distillation described in this invention is as follows:

[0018] Furthermore, the image segmentation model is constructed based on a U-shaped neural network architecture, wherein the image segmentation model includes several decoder layers and an output layer, wherein each encoder layer is connected to the output layer as an intermediate layer branch.

[0019] Furthermore, the loss function used in training the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss is:

[0020]

[0021] Among them, L seg L represents the segmentation loss function. aux_i L represents the branch loss function corresponding to the intermediate layer branch i. penaltyλ represents the alignment loss function between the intermediate layer branch output and the output layer output. aux_i L represents the branch loss function aux_i The coefficient, λ penalty L represents the alignment loss function penalty The coefficient.

[0022] Furthermore, the segmentation loss function L seg for:

[0023]

[0024] Where |G∩P| represents the intersection of the ground truth mask G and the predicted mask P, |G| represents the total number of positive pixels in the ground truth mask, |P| represents the total number of positive pixels in the predicted mask, TP represents true positives, FP represents false positives, and FN represents false negatives.

[0025] Furthermore, the branch loss function L aux_i for:

[0026] L aux_i =α*Dice(S aux_i ,G)+(1-α)*(η f *KL(S aux_i |S finl )+η f *KL(S final |S aux_i ))

[0027] Where Dice represents the splitting result S of branch i. aux_i The Dice coefficient between the real mask G and the ground truth mask G, KL represents the KL divergence, α represents the weight of the segmentation loss function, and η f η represents the coefficient of the positive KL curve. r This represents the coefficient of the reverse KL curve.

[0028] Furthermore, the alignment loss function L penalty for:

[0029]

[0030] Among them, Logit aux_i Logit represents the feature output corresponding to the intermediate branch i. final This represents the feature output of the output layer.

[0031] This invention discloses an image segmentation model training system based on bidirectional hierarchical self-distillation, comprising:

[0032] Build modules are used to construct image segmentation models;

[0033] The training module is used to train the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss to obtain the trained image segmentation model.

[0034] A further improvement of the image segmentation model training system based on bidirectional hierarchical self-distillation described in this invention is as follows:

[0035] Furthermore, the image segmentation model is constructed based on a U-shaped neural network architecture, wherein the image segmentation model includes several decoder layers and an output layer, wherein each encoder layer is connected to the output layer as an intermediate layer branch.

[0036] This invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the image segmentation model training method based on bidirectional hierarchical self-distillation.

[0037] The present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the image segmentation model training method based on bidirectional hierarchical self-distillation.

[0038] The present invention has the following beneficial effects:

[0039] The image segmentation model training method and related apparatus based on bidirectional hierarchical self-distillation described in this invention, in specific operation, uses the Dice loss function and KL divergence bidirectional distillation loss to train the image segmentation model, thereby obtaining the trained image segmentation model. In this way, through bidirectional distillation loss, the low-level structural information can guide the high-level features in reverse, enriching the knowledge transfer path and improving the accuracy of the trained image segmentation model, which is highly practical. Attached Figure Description

[0040] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is an architecture diagram of the image segmentation model in this invention;

[0043] Figure 3 This is a system structure diagram of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0046] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0047] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0048] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0049] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0051] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0052] Example 1

[0053] refer to Figure 1 The image segmentation model training method based on bidirectional hierarchical self-distillation described in this invention includes the following steps:

[0054] 1) Construct an image segmentation model;

[0055] The image segmentation model is built on a U-shaped neural network architecture. The image segmentation model mainly consists of an encoder, a decoder, and the jump links between them. Unlike the commonly used U-shaped architecture, it achieves bidirectional supervision by introducing the output of multiple decoder layers into the output layer.

[0056] For details, please refer to Figure 2The image segmentation model includes a first encoder layer, a second encoder layer, a third encoder layer, a fourth encoder layer, a fifth encoder layer, a first decoder layer, a second decoder layer, a third decoder layer, a fourth decoder layer, a fifth decoder layer, and an output layer. The output of the first encoder layer is connected to the input of the second encoder layer; the output of the second encoder layer is connected to the inputs of the third encoder layer and the fourth decoder layer; the output of the third encoder layer is connected to the inputs of the fourth encoder layer and the third decoder layer; the output of the fourth encoder layer is connected to the input of the fifth encoder layer and the second decoder layer; the output of the fifth decoder layer is connected to the input of the first decoder layer; the output of the first decoder layer is connected to the output layer and the input of the second decoder layer; the output of the second decoder layer is connected to the output layer and the input of the third decoder layer; the output of the third decoder layer is connected to the output layer and the input of the fourth decoder layer; the output of the fourth decoder layer is connected to the input of the fifth decoder layer and the output layer; and the output of the fifth decoder layer is connected to the output layer. The first, second, third, fourth, and fifth decoder layers serve as intermediate layer branches.

[0057] 2) Train the image segmentation model;

[0058] refer to Figure 1 It should be noted that this invention uses the Dice loss function and KL divergence bidirectional distillation loss to train each intermediate layer, and the output layer outputs S. final Treating it as a teacher model, each decoder layer outputs S aux_i The output of the intermediate layer branch is regarded as the student model, in which the teacher model also learns boundary or structural information from the student model.

[0059] Furthermore, the loss function used in the training process of the image segmentation model is:

[0060]

[0061] Among them, L seg L represents the segmentation loss function. aux_i L represents the branch loss function corresponding to the intermediate layer branch i. penalty λ represents the alignment loss function between the intermediate layer branch output and the output layer output. aux_i L represents the branch loss function aux_i The coefficient, denoted as the distillation intensity coefficient, λ. penalty L represents the alignment loss function penalty The coefficient.

[0062] Specifically, the segmentation loss function L segDice Loss is used to measure the overlap between two sets (typically the ground truth segmentation mask G and the predicted segmentation mask P), specifically:

[0063]

[0064] Where |G∩P| represents the intersection of the ground truth mask G and the predicted mask P (in segmentation, it is usually the number of correctly predicted positive pixels, i.e., True Positives, TP); |G| represents the total number of positive pixels in the ground truth mask; |P| represents the total number of positive pixels in the predicted mask; TP represents true positives (correctly predicted foreground pixels); FP represents false positives (background pixels incorrectly predicted as foreground); and FN represents false negatives (foreground pixels incorrectly predicted as background).

[0065] The branch loss function L aux_i The expression is:

[0066] L aux_i =α*Dice(S aux_i ,G)+(1-α)*(η f *KL(S aux_i |S final )+η r *KL(S final |S aux_i ))

[0067] Where Dice represents the splitting result S of branch i. aux_i The Dice Loss between the real mask G and the actual mask G aims to approximate G; KL represents the KL divergence, which learns the output of the "teacher model," and the sum of the two KL divergences is the distillation loss function; α represents the weight of the segmentation loss function, used to balance the segmentation loss function and the distillation loss function; η f The coefficient representing the positive KL curve is denoted as the distillation weight; η r The coefficients representing the reverse KL divergence are denoted as the distillation weights, where the forward KL divergence function is defined as follows:

[0068] It should be noted that the branch loss function L aux_i It has a two-way supervision mechanism, in which the positive KL makes the output S of the intermediate layer branch... aux_i The output S of the output layer final The supervision, and simultaneously its symmetrical opposite, KL makes the output layer's output S final The output S of the intermediate layer branch aux_i Supervision. η f With η rThe settings are shown in Table 1. Since the forward and reverse KL are affected by the number of training rounds, the proportion of reverse KL is set to be larger in the first 25% of rounds, which means that the final output is made closer to the distribution of the intermediate layer branch outputs. The same proportion is set between 25% and 75% of rounds to allow the two to learn from each other. In the final stage III, the proportion of forward KL is made larger, so that the intermediate layer branch outputs are closer to the final output.

[0069] Table 1

[0070]

[0071] The alignment loss function L penalty The expression is:

[0072]

[0073] Among them, Logit aux_i This represents the feature output corresponding to the intermediate branch i; Logit final This represents the feature output of the output layer.

[0074] Distillation intensity coefficient λ aux_i Set to: λ aux_1 =0; λ aux_2 =0.3; λ aux_3 =0.5; λ aux_4 =0.2, here, the contribution of the lowest layer (first decoder layer) is discarded. Low-resolution features (second decoder layer) can capture the global structure, but require moderate weights to avoid overemphasizing coarse patterns; medium-resolution features (third decoder layer) aim to balance global context and local details, so they require higher weights; high-resolution features (fourth decoder layer) will closely match the final output, and only slight adjustments are needed to refine the details, so a smaller weight is set.

[0075] The coefficient λ of the alignment loss function penalty Set to a small constant; in this embodiment, let λ penalty =0.04.

[0076] Example 2

[0077] refer to Figure 3 The image segmentation model training system based on bidirectional hierarchical self-distillation of the present invention includes:

[0078] Build modules are used to construct image segmentation models;

[0079] The training module is used to train the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss to obtain the trained image segmentation model.

[0080] In this embodiment, the image segmentation model is constructed based on a U-shaped neural network architecture. The image segmentation model includes several decoder layers and an output layer, wherein each encoder layer is connected to the output layer as an intermediate branch.

[0081] In this embodiment, the loss function used in training the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss is:

[0082]

[0083] Among them, L seg L represents the segmentation loss function. aux_i L represents the branch loss function corresponding to the intermediate layer branch i. penalty λ represents the alignment loss function between the intermediate layer branch output and the output layer output. aux_i L represents the branch loss function aux_i The coefficient, λ penalty L represents the alignment loss function penalty The coefficient.

[0084] In this embodiment, the segmentation loss function L seg for:

[0085]

[0086] Where |G∩P| represents the intersection of the ground truth mask G and the predicted mask P, |G| represents the total number of positive pixels in the ground truth mask, |P| represents the total number of positive pixels in the predicted mask, TP represents true positives, FP represents false positives, and FN represents false negatives.

[0087] In this embodiment, the branch loss function L aux_i for:

[0088] L aux_i =α*Dice(S aux_i ,G)+(1-α)*(η f *KL(S aux_i |S final )+η r *KL(S final |S aux_i ))

[0089] Where Dice represents the splitting result S of branch i. aux_i The Dice coefficient between the real mask G and the ground truth mask G, KL represents the KL divergence, α represents the weight of the segmentation loss function, and η f η represents the coefficient of the positive KL curve. r This represents the coefficient of the reverse KL curve.

[0090] In this embodiment, the alignment loss function L penalty for:

[0091]

[0092] Among them, Logit aux_i Logit represents the feature output corresponding to the intermediate branch i. final This represents the feature output of the output layer.

[0093] The module division in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in each embodiment of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0094] Example 3

[0095] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a bidirectional hierarchical self-distillation-based image segmentation model training method, for example, including: constructing an image segmentation model; training the image segmentation model using the Dice loss function and KL divergence bidirectional distillation loss to obtain a trained image segmentation model. The memory may include main memory, such as high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which may be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc., and the bus may be divided into an address bus, a data bus, a control bus, etc. The memory is used to store the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0096] Example 4

[0097] A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the image segmentation model training method based on bidirectional hierarchical self-distillation. For example, the steps include: constructing an image segmentation model; training the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss to obtain a trained image segmentation model. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory and / or cache memory, etc. The non-volatile memory may include read-only memory, hard disk, flash memory, optical disk, magnetic disk, etc.

[0098] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0099] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and disclosure of the invention. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0103] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0104] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A training method for an image segmentation model based on bidirectional hierarchical self-distillation, characterized in that, include: Construct an image segmentation model; The image segmentation model is trained using the Dice loss function and the KL divergence bidirectional distillation loss to obtain the trained image segmentation model.

2. The image segmentation model training method based on bidirectional hierarchical self-distillation according to claim 1, characterized in that, The image segmentation model is built on a U-shaped neural network architecture, which includes several decoder layers and an output layer. Each encoder layer is connected to the output layer as an intermediate branch.

3. The image segmentation model training method based on bidirectional hierarchical self-distillation according to claim 2, characterized in that, The loss function used in training the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss is: Among them, L seg L represents the segmentation loss function. aux_i L represents the branch loss function corresponding to the intermediate layer branch i. penalty λ represents the alignment loss function between the intermediate layer branch output and the output layer output. aux_i L represents the branch loss function aux_i The coefficient, λ penalty L represents the alignment loss function penalty The coefficient.

4. The image segmentation model training method based on bidirectional hierarchical self-distillation according to claim 3, characterized in that, The segmentation loss function L seg for: Where |G∩P| represents the intersection of the ground truth mask G and the predicted mask P, |G| represents the total number of positive pixels in the ground truth mask, |P| represents the total number of positive pixels in the predicted mask, TP represents true positives, FP represents false positives, and FN represents false negatives.

5. The image segmentation model training method based on bidirectional hierarchical self-distillation according to claim 3, characterized in that, The branch loss function L aux_i for: L aux_i =α*Dice(S aux_i ,G)+(1-a) *(η f *KL(S aux_i |S final )+η r *KL(S final |S aux_i )) Where Dice represents the splitting result S of branch i. aux_i The Dice coefficient between the real mask G and the ground truth mask G, KL represents the KL divergence, α represents the weight of the segmentation loss function, and η f η represents the coefficient of the positive KL curve. r This represents the coefficient of the reverse KL curve.

6. The image segmentation model training method based on bidirectional hierarchical self-distillation according to claim 3, characterized in that, The alignment loss function L penalty for: Logit aux_i Logit represents the feature output corresponding to the intermediate branch i. final This represents the feature output of the output layer.

7. A training system for an image segmentation model based on bidirectional hierarchical self-distillation, characterized in that, include: Build modules are used to construct image segmentation models; The training module is used to train the image segmentation model using the Dice loss function and the KL divergence bidirectional distillation loss to obtain the trained image segmentation model.

8. The image segmentation model training system based on bidirectional hierarchical self-distillation according to claim 7, characterized in that, The image segmentation model is built on a U-shaped neural network architecture, which includes several decoder layers and an output layer. Each encoder layer is connected to the output layer as an intermediate branch.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the image segmentation model training method based on bidirectional hierarchical self-distillation as described in any one of claims 1-6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the image segmentation model training method based on bidirectional hierarchical self-distillation as described in any one of claims 1-6.

Images