A medical image segmentation method based on multi-expert perception reward and group relative strategy optimization

Abstract

This invention discloses a medical image segmentation method based on multi-expert perceptual reward and group relative policy optimization, aiming to solve the technical pain points of expert annotation discrepancies, low edge localization accuracy, and high memory consumption of traditional reinforcement learning in lung nodule CT image segmentation. This method first constructs an uncertainty-aware field map based on multiple expert-annotated masks, and integrates the MedSAM2 model into the Seg-R1 reinforcement learning framework as the policy network. The GRPO algorithm is used for parallel sampling to generate candidate segmentation variant groups. The relative reward within each group is calculated through a composite reward function that couples weighted Dice and boundary penalties, and parameter updates are completed using a policy pruning mechanism. Finally, an accurate lesion segmentation mask is output. This invention does not require an independent value network, has stronger hardware adaptability, and experimentally measured Dice coefficients reach 0.8318, with HD95 reduced to 3.84 pixels, effectively suppressing false positives and improving edge segmentation accuracy, providing an efficient solution for accurate clinical segmentation of lung nodules.

Description

Technical Field

[0001] This invention relates to a medical image segmentation method based on multi-expert perception reward and group relative strategy optimization. Background Technology

[0002] In clinical diagnosis and treatment, computed tomography (CT) imaging is the primary means of detecting lung lesions. However, pulmonary nodules are typically small in size, have complex anatomical backgrounds, and diverse morphologies. They are also often adhered to surrounding blood vessels, pleura, and other tissues, resulting in low contrast, extremely blurred boundaries, and difficulty in defining edge features in images. This poses a significant technical challenge to accurate automated segmentation.

[0003] In practical applications of medical image analysis, the gold standard for lesion definition faces significant interoperability variability. Due to the influence of clinical experience, cognitive biases, and imaging equipment noise, different radiologists find it difficult to reach a complete consensus on the boundary delineation of the same lesion. This subjectivity leads to widespread consistency discrepancies in medical annotation data: the core area of ​​the lesion usually has high annotation consistency, while the peripheral area is filled with a large amount of uncertainty noise.

[0004] Currently, mainstream deep learning segmentation models, including medical-specific segmentation models such as Convolutional Neural Networks (CNN), Transformer, and MedSAM, mostly employ deterministic supervised learning paradigms. During training, expert annotations are treated as absolute truth, failing to effectively model the knowledge consistency and subjective noise behind the annotations. When faced with edge regions exhibiting strong discrepancies, these models are prone to overfitting or prediction oscillations, resulting in a lack of anatomically coherent segmentation results. This is particularly problematic when dealing with lesion edges with complex geometric features, where sub-millimeter level precision alignment is unattainable.

[0005] To introduce deeper decision-making logic, some studies have applied reinforcement learning to segmentation tasks, endowing models with reasoning capabilities by modeling mask generation as a sequential decision-making process. However, traditional reinforcement learning frameworks face extremely high computational overhead and memory bottlenecks in the field of medical imaging. They typically require maintaining independent value networks to evaluate the quality of segmentation actions. This makes the hardware computing power requirements for processing high-resolution 3D medical images exceed the capabilities of most clinical application environments, greatly limiting the deployment and popularization of reinforcement learning technology in primary healthcare facilities.

[0006] Furthermore, existing segmentation evaluation systems often overemphasize region overlap indicators while neglecting boundary distance error, a physical parameter of greater concern to clinicians. In complex anatomical contexts, even predictions with high overlap may still exhibit significant physical offsets from the actual anatomical boundaries, which is unacceptable in applications such as precision surgical navigation or lesion growth monitoring. Summary of the Invention

[0007] This invention provides a medical image segmentation method based on multi-expert perception reward and group relative strategy optimization to address the problems existing in the prior art.

[0008] The technical solutions adopted in this invention are as follows:

[0009] A medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization includes the following steps:

[0010] The process involves acquiring a medical image to be processed and a bounding box used to indicate the target lesion region, acquiring multiple independent expert annotation masks of the medical image to be processed, constructing an uncertainty-aware field map based on the pixel-level superposition result of the multiple expert annotation masks, and determining the logical union of the multiple expert annotations as the true value. The uncertainty-aware field map is used to characterize the expert annotation confidence weight of each pixel position.

[0011] The mask decoding module of the MedSAM2 model is integrated into the Seg-R1 reinforcement learning framework and defined as a policy network. The policy network receives the medical image to be processed and the bounding box, outputs the pixel-level probability distribution, and generates candidate segmentation masks by sampling under the temperature coefficient constraint.

[0012] The policy network is trained using a group relative policy optimization algorithm: G candidate segmentation masks are generated by parallel sampling of the same medical image and bounding box to form a candidate segmentation variant group. Based on the uncertainty perception field map, the reward value of each candidate segmentation mask in the variant group is calculated by the multi-expert consensus perception reward function. The relative advantage function in the group is obtained by normalization according to the mean and standard deviation of the reward value in the group. The policy network parameters are updated by combining the policy probability ratio pruning mechanism.

[0013] The trained policy network receives the medical images and bounding boxes to be processed and outputs the final lesion segmentation mask.

[0014] Furthermore, an uncertainty-aware field map is constructed based on the pixel-level overlay results of multiple expert-annotated masks, specifically including:

[0015] For each pixel location, count the number of experts who marked it as a lesion, and assign weights based on the number of expert approvals:

[0016] All four experts marked the pixels as lesions with a weight of 1.0;

[0017] The pixels labeled as lesions by the three experts had a weight of 0.75.

[0018] The pixels labeled as lesions by the two experts had a weight of 0.5.

[0019] The pixel weight marked as a lesion by one expert is 0.25;

[0020] Pixels not labeled as lesions by experts have a weight of 0.

[0021] Furthermore, the multi-expert consensus perception reward function is a composite reward function:

[0022] ,

[0023] in, The Dice similarity coefficient is weighted based on the uncertainty-aware field graph. The penalty is 95% of the distance from Hausdorf to the boundary. , These are preset weighting coefficients.

[0024] Furthermore, the weighted Dice reward is calculated as follows:

[0025] ,

[0026] in, For uncertainty-aware field diagrams. For candidate segmentation masks, Label the true values ​​for experts. This is the numerical smoothing constant.

[0027] Furthermore, the 95% Hausdorf distance boundary penalty is specifically as follows:

[0028] ,

[0029] in, For the first The 95% Hausdorff distance between each candidate segmentation mask and the ground truth value labeled by multiple experts For the first One candidate segmentation mask.

[0030] Furthermore, the sampling process controls the randomness of sampling through a temperature coefficient, which is used to adjust the balance between exploration and convergence of pixel-level probability distribution.

[0031] Furthermore, the number of candidate segmentation variant groups is G=8.

[0032] Furthermore, the medical image to be processed is a CT image of a lung nodule, and the target lesion is a lung nodule.

[0033] Furthermore, the pruning threshold of the strategy probability ratio pruning mechanism. Limit the policy probability ratio to Within the range.

[0034] Furthermore, the formula for calculating the within-group relative advantage function is as follows:

[0035] ,

[0036] in, For the first The intra-group relative dominance function values ​​of each candidate segmentation mask. ) is the first The reward value corresponding to each candidate segmentation mask. Within the candidate segmentation variant group The average reward value of each mask, Within the candidate segmentation variant group The standard deviation of the reward value of each mask. This is the numerical smoothing constant.

[0037] The present invention has the following beneficial effects:

[0038] (1) By constructing an uncertainty-aware field map, the independent annotations of four radiology experts are transformed into pixel-level confidence weights, enabling the model to prioritize fitting the core regions where experts have a high degree of consensus during training, and to perform robust learning on the marginal regions where there are disagreements. Compared with the traditional supervised learning approach of treating a single annotation as an absolute truth, this invention can objectively reflect the subjective differences in medical annotations and reduce model overfitting or prediction oscillations caused by disagreements among experts.

[0039] (2) Experimental results on the lung nodule dataset show that the average Dice coefficient of the present invention reaches 0.8318, the IoU crossover ratio reaches 0.7120, and the positive prediction rate (PPV) reaches 0.9015. This indicates that the present invention effectively reduces false positive predictions and reduces the situation where normal tissue is misjudged as lesions while maintaining complete coverage of the main lesion area.

[0040] (3) By introducing a composite reward function that combines a weighted Dice reward based on an uncertainty-aware field map with a Hausdorff distance boundary penalty, the model has higher sensitivity to subtle features such as spiculation and lobulation at the edge of lung nodules. Experimental data show that the 95% Hausdorff distance (HD95) decreased from 14.22 pixels to 3.84 pixels, significantly reducing the edge localization error and helping clinicians to more accurately assess the invasive characteristics of nodules.

[0041] (4) The Group Relative Policy Optimization (GRPO) algorithm is adopted, which estimates the advantage function by the relative reward value of samples within the group, without the need to maintain an independent value network model. Compared with traditional reinforcement learning methods such as Proximal Policy Optimization (PPO) that require additional value networks, this invention has lower memory usage, more stable training process, and is easier to deploy in conventional clinical hardware environments when processing high-resolution medical images.

[0042] (5) The uncertainty perception field map provides a visualized heat map of expert consensus, giving the model's decision-making process clear clinical basis. The model's conservative segmentation strategy in the controversial border region is consistent with the cognitive logic of radiology experts, and the output segmentation mask is more in line with clinical diagnostic experience at the anatomical level. Attached Figure Description

[0043] Figure 1 This is a flowchart of the present invention.

[0044] Figure 2 This is a schematic diagram of the framework of the present invention.

[0045] Figure 3 A comparison diagram of the global lung window, prompt box, expert annotation mask, and segmentation mask of the present invention.

[0046] Figure 4 To optimize the comparison image for segmentation. Detailed Implementation

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] like Figure 1 This invention discloses a medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization, comprising the following steps:

[0049] S1. Data preprocessing: Construct an uncertainty-aware field map reflecting the label confidence level through pixel-level overlay and normalization calculation;

[0050] S2. Introduce the Seg-R1 reinforcement learning framework, initialize the policy network architecture based on MedSAM2, and generate target bounding boxes as geometric cues based on expert-annotated regions to guide the model in locating the lesions to be segmented.

[0051] S3. Employ the GRPO algorithm to generate parallel samples from the same input image. A group of different candidate segmentation results;

[0052] S4. Design a composite reward function that includes an expert consensus perception term and a boundary sensitivity term. Update the model parameters by calculating the relative advantage function within the group to optimize the model's fitting accuracy to complex boundaries.

[0053] S5. Use overlap and boundary distance metrics to verify the accuracy of the output mask and ensure anatomical boundary alignment.

[0054] This embodiment is based on the internationally recognized authoritative dataset for lung nodule research, LIDC-IDRI. This dataset contains chest CT scan images independently annotated by several senior radiologists. After rigorous denoising and quality review, 1843 typical lung nodule samples were selected as the experimental basis data. To adapt to the input requirements of the MedSAM2 model and improve computational efficiency, the CT values ​​of the original DICOM format images were truncated using Hounsfield Unit (HU), with the range set to [-1000, 400] to enhance the contrast between nodules and background tissue. At the same time, resampling was used to eliminate slice thickness differences. The regions of interest of all lung nodules were uniformly scaled to 1024×1024 pixel resolution using a bilinear interpolation algorithm to ensure scale consistency when the model processes lesions of different sizes. Finally, the processed image matrix and the corresponding mask matrix were encapsulated into a binary .npy format file to retain high-precision floating-point information and improve the disk I / O read speed of model training.

[0055] Combination Figure 2 The method of the present invention will be further described below.

[0056] Extract independent segmentation masks from four radiologists for the same CT slice image. , , , By pixel-level accumulation of all expert masks, a multi-expert consensus heatmap is generated. The calculation formula is as follows:

[0057] ,

[0058] In pixels The numerical values ​​reflect the frequency of expert consensus that the location belongs to a nodule region; based on multi-expert consensus heatmaps Constructing an uncertainty-aware field map Differential confidence weights are assigned based on different consistency frequencies, and the specific mapping rules are as follows:

[0059] ,

[0060] The weight field is used to quantify the confidence level of the label. The correspondence between the expert consensus level and the weight coefficient is shown in Table 1 below. It is then used as a weighting factor in the reward function to guide the model training to prioritize fitting regions with high expert consensus and to perform robust learning on dispute edges.

[0061] Table 1 Mapping Table of Expert Consensus Levels and Weighting Coefficients

[0062]

[0063] By introducing the Seg-R1 reinforcement learning framework, the mask decoding module of MedSAM2 is directly defined as the policy network. The policy network is no longer responsible for generating and predicting geometric cue coordinates. Instead, it uses the feature perception capability embedded in the MedSAM2 pre-trained weights to obtain the initial bounding box geometric cue generated based on the expert-annotated region in the early stage of inference. It directly performs Markov decision on the dense pixel-level probability distribution and outputs a mask, eliminating the cascading propagation of segmentation errors caused by small deviations in the network's predicted geometric cue.

[0064] The segmentation task is modeled as a Markov decision process, given the image state. and geometric guidance prompts Policy Network First, a continuous pixel-level probability distribution map is output. Then, under the constraint of temperature coefficient, a specific discretized segmentation mask is generated from this probability distribution as an action. The strategy sampling formula is expressed as:

[0065] ,

[0066] in, Indicates the state and prompt words Under these conditions, the action probability distribution output by the policy network and the temperature coefficient are used to adjust the randomness of the sampling process, balance the exploration space and convergence efficiency of the model, and ensure that the candidate segmentation variants generated by sampling can both identify complex edge features and stably converge to the multi-expert consensus boundary.

[0067] A group-relative policy optimization algorithm is used to train the policy network. For each input 1024×1024 pixel lung nodule image, the random sampling capability of the policy network is utilized to generate images in parallel under temperature coefficient constraints. A set of candidate segmentation masks constitutes a candidate segmentation variant group. This provides a statistical benchmark for calculating the advantage function.

[0068] Based on the uncertainty-aware field graph, a multi-expert consensus-aware composite reward function is designed to calculate the reward value of each candidate segmentation mask within the variant group. The formula for calculating the composite reward function is as follows:

[0069] ,

[0070] in This refers to the Dice reward weighted by the uncertainty-aware field graph. The penalty is 95% of the distance from Hausdorf to the boundary. , To preset the weighting coefficients, this embodiment sets... , .

[0071] The specific calculation method for weighted Dice rewards is as follows:

[0072] ,

[0073] in For uncertainty-aware field diagrams. For the first One candidate segmentation mask, Label the logical union truth values ​​for multiple experts. This is a numerical smoothing constant used to prevent the denominator from being zero during calculation, ensuring the numerical stability of backpropagation and gradient updates in the policy network; the 95% Hausdorff distance boundary penalty term is specifically:

[0074] ,

[0075] in For the first The 95% Hausdorff distance between each candidate segmentation mask and the ground truth labeled by multiple experts is used to penalize the physical offset between the predicted boundary and the true boundary through a negative feedback mechanism, thereby optimizing the model's ability to capture subtle features such as spiculation and lobulation of lung nodules.

[0076] Based on the reward values ​​of each candidate segmentation mask within the variant group, the relative advantage function within the group is calculated to achieve adaptive updates of the policy network parameters without relying on the independent value network model. The formula for calculating the relative advantage function within the group is as follows:

[0077] ,

[0078] in, For the first The intra-group relative dominance function values ​​of each candidate segmentation mask. For the first The reward value corresponding to each candidate segmentation mask. Within the candidate segmentation variant group The average reward value of each mask, Within the candidate segmentation variant group The standard deviation of the reward value of each mask. This is the numerical smoothing constant.

[0079] The GRPO loss function is calculated and the policy network parameters are updated using the policy probability ratio pruning mechanism. The formulas for calculating the policy probability ratio and the GRPO loss function are as follows:

[0080] , ,

[0081] Among them, the cropping threshold Limit the policy probability ratio to Within the specified range, to prevent excessive policy update fluctuations from causing training oscillations, the AdamW optimizer is used to perform backpropagation on the loss function, with an initial learning rate set to [value missing]. The model training is completed in conjunction with the cosine annealing learning rate scheduling algorithm.

[0082] After training, the inference phase begins. The CT images of the lung nodules to be processed and the bounding boxes indicating the target lesion regions are input into the trained policy network. The policy network outputs a pixel-level probability distribution, which, after thresholding, yields a binarized final lesion segmentation mask. The segmentation results are quantitatively evaluated and anatomical boundary consistency verified using multidimensional indices such as Dice, IoU, PPV, and HD95. The calculation formulas for each index are as follows:

[0083] Dice coefficient calculation formula: ;

[0084] IoU (Intersection over Union) Ratio Calculation Formula: ;

[0085] Formula for calculating positive predictive rate: ;

[0086] in, This represents the number of true positive pixels, i.e., the regions that the model correctly predicted as nodules. This represents the number of false positive pixels, i.e., the background area that the model mistakenly identifies as a nodule; the formula for calculating the one-way Hausdorff distance is:

[0087] ;

[0088] Formula for calculating 95% Hausdorf distance:

[0089] ;

[0090] in, To predict the mask, For a truth value agreed upon by multiple experts, For point set Point set One-way Hausdorf distance, To obtain the 95th percentile value, extreme interference from isolated noise points is eliminated.

[0091] By comprehensively comparing the Dice, IoU, and HD95 metrics, as shown in Table 2, the model effectively solves the feature loss problem caused by edge "smoothing" while ensuring region coverage. This evaluation mechanism ensures that the segmentation mask not only excels in mathematical metrics but also better aligns with radiologists' understanding of anatomical boundaries.

[0092] Table 2 Indicator Evaluation

[0093]

[0094] This experiment measured various performance metrics on 1843 lung nodule samples, and we analyzed the results. The core of this experiment was to verify the synergistic gain effect of the "multi-expert uncertainty-aware reward function" and the "GRPO group relative strategy optimization algorithm" in complex medical image segmentation tasks. By comparing the data performance with the benchmark model, the data after introducing the reinforcement learning framework, and the data after superimposing the innovative algorithm of this invention, a step-like improvement in the model's performance in three dimensions—region overlap, segmentation accuracy, and anatomical boundary fitting accuracy—can be clearly observed.

[0095] In medical image segmentation, the Dice coefficient and Intersection over Union (IoU) are core metrics for evaluating the model's coverage of the lesion region. Experimental data shows that the basic MedSAM2 model has an average Dice coefficient of 0.7385 and an IoU of 0.5854 when processing samples. This result indicates that although MedSAM2 possesses strong basic semantic extraction capabilities, traditional single-batch supervised learning paradigms struggle to fully capture the complex boundary logic of lesions like lung nodules, which exhibit extremely low contrast and severe adhesion to surrounding blood vessels or the pleura.

[0096] When the Seg-R1 reinforcement learning framework was introduced, the model performance was significantly improved, with the Dice coefficient jumping to 0.7951. The essence of this improvement lies in the fact that reinforcement learning transforms the segmentation process from "pixel-level static mapping" to "sequential decision-making actions." The model no longer passively fits labels but searches for policies within the sampling space, thus possessing stronger feature exploration capabilities. Furthermore, after combining the core GRPO algorithm and the expert uncertainty-aware weighting mechanism of this invention, the Dice coefficient finally reached 0.8318, and the IoU improved to 0.7120. This result quantitatively demonstrates that the technical approach of guiding the model to converge towards a multi-expert consensus region by utilizing intra-group relative advantages is extremely effective, enabling the model to achieve extremely high region consistency with a large sample size of 1843.

[0097] In lung CT images, vascular shadows, pleural thickening, or lung parenchymal inflammation often closely resemble lesions, easily leading to over-segmentation by the model. Experimental data shows that the baseline model's PPV is only 0.7512, meaning that approximately one-quarter of the predicted pixels are "false positives." Our proposed solution improves the PPV to 0.9015. This significant improvement in accuracy stems from our clever introduction of an uncertainty-aware field map into the reward function. By classifying the annotations from four experts in the dataset, the model is forced to prioritize fitting high-confidence core regions fully annotated by all four experts during training, while robust regression strategies are applied to controversial edge regions annotated by only one expert. This design not only solves the annotation noise problem caused by expert disagreements but also significantly enhances the model's ability to exclude interfering tissues, ensuring the high rigor of the segmentation results at the anatomical level.

[0098] For pulmonary nodules, the presence of spiculated or lobulated edges is crucial evidence for determining benignity or malignancy. Traditional segmentation algorithms often focus only on overall regional overlap, easily resulting in overly smooth edges that miss these critical clinical details. In this experiment, the 95% Hausdorff distance decreased dramatically from 14.22 pixels in the baseline model to 3.84 pixels. This error reduction is attributed to the boundary distance-sensitive term designed into the reward function of this invention. Unlike Dice, which only considers area, the boundary-sensitive reward directly perceives the physical displacement between the predicted edge and the edge of multi-expert uncertainty. In the adaptive evolution of 1843 nodule samples, the model accurately captured subtle signs of nodule edges. This improvement is not only reflected in the reduction of numerical indicators but also in the high degree of fit between the final segmentation mask and the actual anatomical structure, providing clinicians with more interpretable auxiliary diagnostic evidence.

[0099] from Figure 3 The visualization results allow for a direct observation of the technical advantages of the method described in this invention in lesion segmentation tasks. Comparing the expert-annotated mask in the third column with the predicted mask of this invention in the fourth column reveals that, under the decision-making framework of Seg-R1 reinforcement learning, the model-predicted mask not only closely matches the actual annotations in terms of subject coverage but also demonstrates superior performance in fitting the details of anatomical edges.

[0100] from Figure 4 Comparative analysis reveals that the basic model is susceptible to interference from complex backgrounds when processing lung nodules, resulting in noticeable undersegmentation or voids at the edges. After introducing the Seg-R1 reinforcement learning framework, the model initially demonstrated policy search capabilities, significantly improving the completeness of the segmented regions. The final optimized scheme of this invention performed best, accurately identifying minute burrs and blurred boundaries of the nodules at the pixel level, with segmentation results conforming to anatomical morphology and exhibiting high fidelity.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on multi-expert perception reward and group relative strategy optimization, characterized by: Includes the following steps: The process involves acquiring a medical image to be processed and a bounding box used to indicate the target lesion region, acquiring multiple independent expert annotation masks of the medical image to be processed, constructing an uncertainty-aware field map based on the pixel-level superposition result of the multiple expert annotation masks, and determining the logical union of the multiple expert annotations as the true value. The uncertainty-aware field map is used to characterize the expert annotation confidence weight of each pixel position. The mask decoding module of the MedSAM2 model is integrated into the Seg-R1 reinforcement learning framework and defined as a policy network. The policy network receives the medical image to be processed and the bounding box, outputs the pixel-level probability distribution, and generates candidate segmentation masks by sampling under the temperature coefficient constraint. The policy network is trained using a group relative policy optimization algorithm: G candidate segmentation masks are generated by parallel sampling of the same medical image and bounding box to form a candidate segmentation variant group. Based on the uncertainty perception field map, the reward value of each candidate segmentation mask in the variant group is calculated by the multi-expert consensus perception reward function. The relative advantage function in the group is obtained by normalization according to the mean and standard deviation of the reward value in the group. The policy network parameters are updated by combining the policy probability ratio pruning mechanism. The trained policy network receives the medical images and bounding boxes to be processed and outputs the final lesion segmentation mask.

2. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: An uncertainty-aware field map is constructed based on the pixel-level overlay results of multiple expert-annotated masks, specifically including: For each pixel location, count the number of experts who marked it as a lesion, and assign weights based on the number of expert approvals: All four experts assigned a weight of 1.0 to the pixels marked as lesions. The pixels labeled as lesions by the three experts had a weight of 0.

75. The pixels labeled as lesions by the two experts had a weight of 0.

5. The pixel weight marked as a lesion by one expert is 0.25; Pixels not labeled as lesions by experts have a weight of 0.

3. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: The multi-expert consensus perception reward function is a composite reward function: ， in, The Dice similarity coefficient is weighted based on the uncertainty-aware field graph. The penalty is 95% of the distance from Hausdorf to the boundary. , These are preset weighting coefficients.

4. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 3, characterized in that: The weighted Dice reward is calculated as follows: ， in, For uncertainty-aware field diagrams. For candidate segmentation masks, Label the true values ​​for experts. This is the numerical smoothing constant.

5. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 3, characterized in that: The 95% Hausdorf distance boundary penalty is specifically as follows: ， in, For the first The 95% Hausdorff distance between each candidate segmentation mask and the ground truth value labeled by multiple experts For the first One candidate segmentation mask.

6. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: The sampling process controls the randomness of sampling through a temperature coefficient, which is used to adjust the balance between exploration and convergence of pixel-level probability distribution.

7. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: The number of candidate segmentation variant groups is G=8.

8. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in any one of claims 1 to 7, characterized in that: The medical image to be processed is a CT image of a lung nodule, and the target lesion is a lung nodule.

9. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: Clipping threshold of the probability ratio clipping mechanism Limit the policy probability ratio to Within the range.

10. The medical image segmentation method based on multi-expert perceptual reward and group relative strategy optimization as described in claim 1, characterized in that: The formula for calculating the within-group relative advantage function is: ， in, For the first The intra-group relative dominance function values ​​of each candidate segmentation mask. ) is the first The reward value corresponding to each candidate segmentation mask. Within the candidate segmentation variant group The average reward value of each mask, Within the candidate segmentation variant group The standard deviation of the reward value of each mask. This is the numerical smoothing constant.

Images