A prostate lesion region segmentation method based on a distillation diffusion model

Abstract

The application provides a prostate lesion region segmentation method based on a distillation diffusion model, comprising the following steps: S1, obtaining a prostate medical image and a corresponding true value mask thereof; S2, inputting input data into a teacher diffusion model for training to obtain a teacher diffusion model with dynamic region perception capability; S3, generating a single-step distillation target aligned with a feature space of the teacher model; S4, obtaining a trained student diffusion model; and S5, inputting a prostate medical image to be segmented into the trained student diffusion model to obtain a final prostate lesion region segmentation result. Through a dynamic region attention mechanism and a time alignment distillation strategy, the application not only significantly improves the accuracy of the segmentation result, especially the accurate delineation of the lesion boundary, but also compresses a complex and iterative denoising process into a single-step inference, thereby providing effective technical support for realizing fast, accurate and automatic segmentation of prostate lesions.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more particularly to a method for segmenting prostate lesion regions based on a distillation-diffusion model. Background Technology

[0002] Prostate cancer is one of the most common malignant tumors in men, and its accurate diagnosis and treatment planning heavily rely on the accurate segmentation of the prostate and its lesion areas. Magnetic resonance imaging (MR) and ultrasound imaging are the primary means of screening and diagnosing prostate diseases. However, in medical images, the prostate and its lesion areas often lack clear boundaries and sufficient contrast with surrounding tissues, and their shape, size, and intensity distribution exhibit significant individual variations, making accurate segmentation extremely challenging. With the development of deep learning technology, convolutional neural networks (CNNs) have made significant progress in medical image segmentation. For example, models such as U-Net and its variants automatically segment prostate regions by learning image features. However, these CNN-based deterministic models have inherent limitations: they typically produce a single output for each input, failing to reflect the inherent ambiguity and uncertainty in medical image annotation, such as differences in how different physicians determine boundary regions. This lack of uncertainty representation limits the reference value of model predictions in clinical decision-making.

[0003] In recent years, the diffusion probability model (DPM) has achieved groundbreaking success in the field of image generation. By simulating the stochastic process of noise addition and removal, it can generate diverse, high-quality samples. This characteristic has led to its introduction into the field of medical image segmentation, resulting in a series of DPM-based segmentation methods. These methods generate segmentation masks from random noise through a multi-step denoising process, and can estimate the uncertainty of the segmentation results through multiple samplings, providing richer decision-making information than deterministic models.

[0004] However, existing prostate segmentation methods based on DPM still face two key challenges: (1) In terms of model accuracy, existing methods fail to fully consider the dynamic characteristics of the denoising process and prior knowledge of the prostate's anatomical structure. Specifically, existing techniques typically employ a static conditional network, injecting the same image features into the diffusion model throughout the denoising process. This approach ignores the fact that the model's focus area changes dynamically at different denoising stages: in the early stages of denoising at high noise levels, the model's primary task is to accurately locate the core region of the lesion; while in the later stages of denoising at low noise levels, the model should focus on refining the blurred boundaries of the lesion. Existing methods employ uniformly weighted attention mechanisms or simple feature injection strategies, which cannot model this dynamic priority shift from core localization to boundary refinement, leading to unreliable predictions in blurred boundary regions. (2) In terms of inference efficiency, the inherent iterative sampling mechanism of DPM results in high computational costs, making it difficult to meet the real-time requirements of clinical practice. Standard DPM may require hundreds or even thousands of denoising steps to generate a result, making it unsuitable for real-time scenarios such as surgical navigation and point-of-care diagnosis. To accelerate inference, existing technologies mainly focus on two types of methods: one is to develop efficient numerical solvers to reduce the number of steps to a few dozen, but the iterative nature remains unchanged, and the speed improvement is limited; the other is to use knowledge distillation to transfer knowledge from a large model to a small model. However, existing distillation methods usually require complex multi-stage training, and the final result is still a model that requires multi-step inference. Furthermore, distillation methods aimed at achieving single-step generation, while performing well in the natural image domain, suffer significant performance degradation when directly applied to medical image segmentation due to feature domain differences. That is, the teacher model benefits from the rich multi-scale, time-related features observed during multi-step denoising, while the student model is forced to directly learn complex output mappings from a single time step. This mismatch between input and context makes it difficult for the student model to inherit the robustness of the teacher model.

[0005] Therefore, there is an urgent need in this field for a prostate lesion segmentation scheme that can both simulate the dynamic segmentation priority of the prostate region to improve accuracy and achieve high-speed single-step reasoning through an effective distillation mechanism. Summary of the Invention

[0006] In view of this, the purpose of this invention is to propose a method for segmenting prostate lesion regions based on a distillation-diffusion model, in order to solve the technical problem that existing methods fail to fully consider the dynamic characteristics of the denoising process and prior knowledge of the prostate's anatomical structure.

[0007] The technical means employed in this invention are as follows: A method for segmenting prostate lesion regions based on a distillation-diffusion model includes the following steps: S1. Obtain the medical image of the prostate and its corresponding ground truth mask as input data; S2. Input the input data into the teacher diffusion model for training to obtain a teacher diffusion model with dynamic region perception capability. The teacher diffusion model adaptively adjusts the attention to the core region and boundary region of the lesion according to the time step of the reverse denoising process through its internal sparse temporal region attention module, generating time-aware multi-step intermediate features. S3. Input the time-aware multi-step intermediate features into the time-aligned projection network, and perform feature projection and fusion processing through the time-aligned projection network to generate a single-step distillation target aligned with the feature space of the teacher model. S4. Using the single-step distillation target, the student diffusion model is trained using the time-aligned single-step distillation method, so that the single-step output of the student diffusion model approximates the single-step distillation target, thereby obtaining the trained student diffusion model. S5. Input the prostate medical image to be segmented into the trained student diffusion model, and obtain the final prostate lesion region segmentation result through single-step forward inference.

[0008] Furthermore, in S2, the execution process of the sparse temporal region attention module includes: S21. Extract multi-scale features of the input image using an encoder; S22. Multi-scale features are processed by boundary conditioner, core conditioner and prostate conditioner respectively to generate boundary feature group, core feature group and overall prostate feature group; S23. Based on the output features of the self-attention layer in the overall prostate feature group and the teacher diffusion model, generate a binarized region mask; S24. Guided by the binarized region mask, sparse matrix operations are performed on the boundary feature group and the core feature group to obtain the boundary attention map and the core attention map, respectively. S25. Based on the current denoising time step, perform dynamic temporal weighted fusion of the boundary attention map and the core attention map to generate a time-adaptive attention map, and apply it to the self-attention output features.

[0009] Furthermore, the time-adaptive attention map Calculated using the following formula:

[0010] in, This is the normalized denoising time step, with a value between 0 and 1. For boundary attention maps, This is the core attention graph, where ⊙ represents element-wise multiplication and ⊕ represents element-wise addition.

[0011] Furthermore, the truth labels used by the core conditioner and boundary conditioner are obtained by performing a distance transformation on the prostate truth mask, specifically including: transforming the truth mask... Perform distance transformation and normalization to obtain the centrality mapping graph. Core area tags pass Calculated; Boundary region label pass Calculated.

[0012] Furthermore, the time-aligned projection network is composed of multiple stacked feedforward neural networks. The input of the time-aligned projection network is the intermediate layer features of the student model at a specific time step and the embedding representation of that time step. The output of the time-aligned projection network is the projected features aligned with the corresponding time step features of the teacher model.

[0013] Furthermore, the loss function of the time-aligned single-step distillation method includes: Distillation loss is used to align intermediate features between the teacher and student models; The segmentation loss is used to constrain the consistency between the student model output and the real mask.

[0014] Furthermore, the distillation loss is calculated using a negative cosine similarity loss function.

[0015] Furthermore, when training the student diffusion model in S4, the network weights of the encoder and conditioner are initialized with the weights of the corresponding parts in the trained teacher model.

[0016] The present invention also provides a storage medium comprising a stored program, wherein, when the program is executed, it performs any of the above-described methods for segmenting prostate lesion regions based on a distillation-diffusion model.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes any of the above-described methods for segmenting prostate lesion regions based on a distillation-diffusion model through the computer program.

[0018] Compared with the prior art, the present invention has the following advantages: This invention employs a dynamic region attention mechanism (STRA), which allows the model to mimic the diagnostic logic of experts and rationally allocate attention resources during the generation process. This results in more accurate and coherent segmentation contours on challenging cases (such as those with blurred boundaries and low contrast), with significant improvements in boundary-sensitive metrics such as HD95 and ASD.

[0019] This invention accelerates inference by compressing the diffusion process, which originally required hundreds of iterations, into a single-step inference process through Time-Aligned Single-Step Distillation (TOD). This enables the invention to be applied to clinical scenarios with stringent real-time requirements, such as surgical navigation and point-of-care diagnosis.

[0020] This invention solves the distillation bottleneck of feature domain offset by using a time-aligned projection network, enabling the student model to not only match the final output of the teacher model, but also learn its rich intermediate feature representations. As a result, while achieving a huge speed improvement, the segmentation accuracy can be comparable to the original multi-step teacher model, and even surpass it in some metrics.

[0021] The method of this invention maintains superior performance under different imaging devices and different acquisition protocols, demonstrating strong generalization ability. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is an overall structural diagram of the invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the structure of the sparse temporal region attention module, a core component of the teacher diffusion model of this invention. Figure 4 This is the time-aligned projection network diagram of the present invention; Figure 5 This is a schematic diagram illustrating how the dynamic attention weights of the present invention change with the denoising time step; Figure 6 This is a visual comparison of the method of the present invention with other existing methods in the prostate segmentation task. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 the invention 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 a 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] like Figure 1 and 2 As shown, this invention provides a method for segmenting prostate lesion regions based on a distillation-diffusion model, comprising the following steps: S1. Obtain the medical image of the prostate and its corresponding ground truth mask as input data; S2. Input the input data into the teacher diffusion model for training to obtain a teacher diffusion model with dynamic region perception capability. The teacher diffusion model adaptively adjusts the attention to the core region and boundary region of the lesion according to the time step of the reverse denoising process through its internal sparse temporal region attention module, generating time-aware multi-step intermediate features. S3. Input the time-aware multi-step intermediate features into the time-aligned projection network, and perform feature projection and fusion processing through the time-aligned projection network to generate a single-step distillation target aligned with the feature space of the teacher model. S4. Using the single-step distillation target, the student diffusion model is trained using the time-aligned single-step distillation method, so that the single-step output of the student diffusion model approximates the single-step distillation target, thereby obtaining the trained student diffusion model. S5. Input the prostate medical image to be segmented into the trained student diffusion model, and obtain the final prostate lesion region segmentation result through single-step forward inference.

[0027] It mainly includes two core stages: dynamic region-aware teacher model training and time-aligned single-step knowledge distillation.

[0028] Phase 1: Training of teacher models for dynamic region perception; The goal of this phase is to train a powerful teacher diffusion model. It can dynamically and adaptively focus on different anatomical sub-regions of prostate lesions through the sparse temporal region attention module proposed in this invention during the forward noise addition and reverse noise reduction process.

[0029] Step 1: Decoupling data preprocessing from region labeling; Obtaining a prostate medical image dataset and its corresponding binarized segmentation truth mask .right Perform a distance transformation and normalize to the interval [0,1] to obtain the centrality mapping graph. .based on The core area supervision label is decoupled using formula (1). and border area supervision labels :

[0030] in, This indicates element-wise multiplication. Areas with higher pixel values ​​are closer to the geometric center of the lesion. This places more emphasis on the boundary area between the lesion and the background.

[0031] Step 2: Multi-scale region feature extraction; Input image Input encoder (For example, PVT-B2), extract image features at four levels, denoted as... ,in The lowest level (high resolution) features, This represents the top-level (low-resolution) features. These features are then processed in parallel by three conditionals with different structures to learn the characteristics of different regions. Furthermore, the feature channels described above can be unified. First, a transformation layer is used. (Typically 1x1 convolution, batch normalization, and ReLU activation function) Maps features from all layers to a uniform number of channels. (For example, 64), to prepare for subsequent processing:

[0032] in, and These are the boundary augmentation conditioner and the core augmentation conditioner at the 1st... The initial input features of the layer.

[0033] Step 3: Construct boundary enhancement, core enhancement, and prostate conditioner; The boundary enhancement conditioner employs an equilateral triangular structure, aiming to progressively integrate features from top to bottom, focusing on learning edge texture features. For the first... line, number Column characteristics The calculation formula is as follows:

[0034] in This indicates a channel dimension splicing operation. This indicates a 2x upsampling operation. This represents a basic convolutional block, consisting of a 3x3 convolution, batch normalization, and a ReLU activation function. (Variable) The definition of is:

[0035] The output of the boundary enhancement conditioner is a multi-scale boundary feature set. , recorded as .

[0036] The core augmenting conditioner employs an inverted triangle structure, aiming to abstract features progressively from the bottom up, focusing on learning core semantic features. For the first... line, number Column characteristics The calculation formula is as follows:

[0037] in This indicates element-wise addition. This also represents the basic convolutional block. The variable is defined as:

[0038] The core enhancement conditioner outputs a multi-scale core feature set. , recorded as .

[0039] Prostate conditioner As a feature fusion processor, a simplified FPN structure is employed to integrate the multi-scale features output by the boundary enhancement conditioner and the core enhancement conditioner, generating holistic prostate features to guide the diffusion model. Layer output The calculation formula is:

[0040] The output of the prostate conditioner is the multi-scale overall prostate feature set. , recorded as .

[0041] Step 4: Conditioner supervised training; To guide the three conditioners in learning effective features, decoupled soft labels are used for supervision. Total loss. It is a weighted sum of multiple monitoring signals:

[0042]

[0043] in, , , These represent the binary cross-entropy loss, cross-union ratio loss, and Dice loss, respectively.

[0044] Phase Two: Constructing a sparse temporal region attention mechanism; After obtaining specially enhanced multi-scale region features generated by boundary conditioners, core conditioners, and prostate conditioners, we obtained multi-scale region features. , , Subsequently, the core innovative module of this invention—the sparse temporal region attention module—begins to function. The STRA module is integrated into each Transformer block of the teacher diffusion model, replacing the standard cross-attention layer. Its design goal is to dynamically and efficiently inject the aforementioned region features into the diffusion backbone, and ensure that at different time steps in the denoising process, the model's focus on the region smoothly transitions from core localization to boundary refinement. The sparse temporal region attention module is the core of the teacher model, and its structure is as follows: Figure 3 As shown, the workflow is as follows: Step 1: Identification and sparsification of regional importance; First, STRA needs to determine which spatial locations in the current feature map belong to the prostate-related regions that require focused computation, in order to significantly reduce unnecessary calculations. This process begins by generating a binary mask of region importance using overall prostate features. Multi-scale prostate features are then fused. It is implemented through a lightweight Feature Pyramid Network (FPN) and a PatchEmbedding operation. Transform it into a teacher diffusion network. Self-attention output features of layer Transformer blocks Key / value features with the same spatial dimensions. Compare this feature with... Matrix multiplication is performed, and the sigmoid function is applied to generate a preliminary attention map with values ​​in the range [0,1]. Through the analysis of Set a threshold (e.g., 0.5), binarize it, and you will get the region importance mask. :

[0045] in . This represents matrix multiplication. From this, we obtain the set of spatial location indices that require the most computation. .

[0046] Step 2: Sparse attention calculation; After obtaining the sparse location set Next, the STRA module calculates refined attention maps for the boundary and core regions respectively. This step is crucial for improving the computational efficiency of this module. To reduce computational complexity, only the aforementioned masks... Calculate the boundary attention map within the specified set of sparse locations. and core attention map :

[0047] in, This represents a partial matrix multiplication operation. Subsequently, these calculated weights are determined according to their indices. Fill back and A complete boundary attention map is obtained from a matrix of the same size. and core attention map This operation is recorded as .

[0048] Step 3: Temporal dynamic modulation and feature injection; At time step (Normalized to [0,1]), for and Dynamic weighted fusion is performed to generate the final continuous-time adaptive attention map. :

[0049] in, This represents element-wise multiplication. This indicates element-wise addition. Indicates This refers to the cross-attention operation for attention weights. This is the output of the STRA module, which injects dynamically weighted region features into the diffusion model. This design allows the model to perform well in the early stages of denoising. More emphasis is placed on boundary refinement in the later stages of noise reduction. More focus on core positioning. For example... Figure 5 As shown, the attention weights in this method enable the model to perform well in the early stages of denoising. More emphasis is placed on boundary refinement in the later stages of noise reduction. More focus is placed on core positioning. Next, an autoencoder is used to verify the truth value. Perceptual compression is performed to obtain the latent representation. The forward noise addition process is defined as follows: ,in The noise is Gaussian. (Teacher model) Total training loss for:

[0050] in, This represents the mean square error loss.

[0051] Phase 3: Time-aligned single-step knowledge distillation; While the teacher model boasts excellent segmentation performance, its iterative sampling process (typically requiring hundreds of steps) is computationally expensive. The goal of this phase is to fully transfer the knowledge from the teacher model to a student model, which can complete segmentation with only a single forward propagation step, achieving a significant leap in inference speed. Multi-step knowledge distillation into a student model This enables it to perform high-speed single-step inference without sacrificing performance.

[0052] Step 1: Construct the model initialization and time-aligned projection network; Student model structure: Student model The network structure is basically the same as the teacher model, but the STRA module in its diffusion network is replaced with a standard cross-attention layer to reduce computational complexity during inference. Weight initialization: To accelerate convergence and provide a good starting point, the weights of the encoder and three conditioners from the teacher model trained in the first stage are directly copied to the corresponding parts of the student model. This gives the student model strong feature extraction capabilities from the beginning. Temporal alignment projection network: To address the feature domain gap between teacher and student models caused by different input contexts (teachers observe multiple steps, students only observe a single step), this invention introduces a lightweight temporal alignment projection network. The temporal alignment projection network is constructed as follows: Figure 4 As shown. The TPN consists of stacked feedforward network blocks and a time conditioner. Composition. For the student model in the first... Intermediate features of the layer TPN output for:

[0053] The core function of TPN is feature alignment: it projects the immature features generated by the student model in single-step inference into a feature space that is compatible with the mature features generated by the teacher model after being processed by the complex STRA module at the corresponding time step.

[0054] Step 2: Distillation training and loss function; At this stage, the teacher model is frozen. All parameters, trained only on the student model And TPN. The loss function is carefully designed in two parts to ensure effective knowledge transfer. All parameters are used to train only the student model. And TPN. The loss function is carefully designed in two parts to ensure the effective transfer of knowledge. Total loss from distillation. It consists of two parts

[0055] in , The negative cosine similarity loss is employed, with the loss occurring across multiple network layers (e.g., 5 encoder layers and 1 intermediate bottleneck layer, totaling...). The difference between the student's features projected by TPN and the teacher's STRA output features is calculated on the layer. To segment the loss, the mean squared error between the student model's single-step output and the true label latent variable is calculated. It is a weighted hyperparameter that balances the two losses. By combining this multi-level feature alignment with the final output constraint, the student model is forced not only to know what (the final result), but also why (the internal feature evolution process), thus enabling it to reproduce the high performance of the teacher model in single-step inference.

[0056] After training, during the inference phase, the time-aligned projection network completes its task and is discarded. For a prostate image to be segmented... Sample a noise tensor from a standard Gaussian distribution Subsequently, and Input into the trained student model In the middle, a single forward propagation is performed. The model's output is then passed through the decoder. After upsampling, the final high-precision segmentation prediction map is obtained. This process achieves a speedup of hundreds of times compared to the original diffusion model, making it potential for application in real-time clinical scenarios. For example... Figure 6 As shown, this method not only outperforms the other two diffusion model methods, but also outperforms some other state-of-the-art methods.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for segmenting prostate lesion regions based on a distillation-diffusion model, characterized in that, Includes the following steps: S1. Obtain the medical image of the prostate and its corresponding ground truth mask as input data; S2. Input the input data into the teacher diffusion model for training to obtain a teacher diffusion model with dynamic region perception capability. The teacher diffusion model adaptively adjusts the attention to the core region and boundary region of the lesion according to the time step of the reverse denoising process through its internal sparse temporal region attention module, generating time-aware multi-step intermediate features. S3. Input the time-aware multi-step intermediate features into the time-aligned projection network, and perform feature projection and fusion processing through the time-aligned projection network to generate a single-step distillation target aligned with the feature space of the teacher model. S4. Using the single-step distillation target, the student diffusion model is trained using the time-aligned single-step distillation method, so that the single-step output of the student diffusion model approximates the single-step distillation target, thereby obtaining the trained student diffusion model. S5. Input the prostate medical image to be segmented into the trained student diffusion model, and obtain the final prostate lesion region segmentation result through single-step forward inference.

2. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 1, characterized in that, In S2, the execution process of the sparse temporal region attention module includes: S21. Extract multi-scale features of the input image using an encoder; S22. Multi-scale features are processed by boundary conditioner, core conditioner and prostate conditioner respectively to generate boundary feature group, core feature group and overall prostate feature group; S23. Based on the output features of the self-attention layer in the overall prostate feature group and the teacher diffusion model, generate a binarized region mask; S24. Guided by the binarized region mask, sparse matrix operations are performed on the boundary feature group and the core feature group to obtain the boundary attention map and the core attention map, respectively. S25. Based on the current denoising time step, perform dynamic temporal weighted fusion of the boundary attention map and the core attention map to generate a time-adaptive attention map, and apply it to the self-attention output features.

3. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 2, characterized in that, The time-adaptive attention graph Calculated using the following formula: in, This is the normalized denoising time step, with a value between 0 and 1. For boundary attention maps, This is the core attention graph, where ⊙ represents element-wise multiplication and ⊕ represents element-wise addition.

4. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 2, characterized in that, The truth labels used by the core conditioner and boundary conditioner are obtained by performing a distance transform on the prostate truth mask, specifically including: transforming the truth mask... Perform distance transformation and normalization to obtain the centrality mapping graph. Core area tags pass Calculated; Boundary region label pass Calculated.

5. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 1, characterized in that, The time-aligned projection network is composed of multiple feedforward neural networks stacked together. The input of the time-aligned projection network is the intermediate layer features of the student model at a specific time step and the embedding representation of that time step. The output of the time-aligned projection network is the projected features aligned with the corresponding time step features of the teacher model.

6. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 1, characterized in that, The loss function of the time-aligned single-step distillation method includes: Distillation loss is used to align intermediate features between the teacher and student models; The segmentation loss is used to constrain the consistency between the student model output and the real mask.

7. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 6, characterized in that, The distillation loss is calculated using the negative cosine similarity loss function.

8. The method for segmenting prostate lesion regions based on a distillation-diffusion model according to claim 1, characterized in that, When training the student diffusion model in S4, the network weights of the encoder and conditioner are initialized with the weights of the corresponding parts in the trained teacher model.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein when the program is executed, it performs the prostate lesion region segmentation method based on the distillation diffusion model as described in any one of claims 1 to 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the method for segmenting prostate lesion regions based on the distillation diffusion model as described in any one of claims 1 to 8 through the computer program.

Images