A pathological image segmentation method based on fractional order memory enhancement and differential injection

Abstract

This invention discloses a pathological image segmentation method based on fractional-order memory enhancement and differential injection, belonging to the fields of intelligent medical image processing and computer vision technology. It addresses the problems of high noise in shallow features, insufficient utilization of cross-layer feature memory, blurred boundaries, and difficulty in separating adhered targets in existing pathological image segmentation methods. The method inputs the pathological image into a fractional-order memory enhancement encoding-decoding segmentation network. A foreground probability map is obtained through staged differential injection using shared fractional-order pre-denoising, fractional-order memory skip connections, and shared fractional-order differential responses. Counterfactual candidate results such as preservation, repair, splitting, and suppression are then constructed and selected based on joint energy evaluation. A segmentation mask is output to improve boundary preservation, separation of adhered targets, hole repair, and false detection suppression capabilities. This method can be used for segmenting pathological targets such as glands and cell nuclei.

Description

Technical Field

[0001] This invention relates to the fields of intelligent medical image processing and computer vision technology, specifically to a pathological image segmentation method based on fractional-order memory enhancement, shared fractional-order pre-denoising, shared fractional-order differential response, and counterfactual post-processing selection. Background Technology

[0002] With the development of digital pathology scanning equipment and medical image processing technology, automatic pathology image segmentation has become an important technical means to assist doctors in histological analysis, cell counting, glandular structure identification, and lesion boundary localization. Among existing pathology image segmentation methods, deep learning methods based on encoder-decoder structures, especially U-Net and its improved methods, are widely used because they combine multi-scale semantic modeling with spatial resolution restoration capabilities.

[0003] However, pathological images exhibit more complex texture distributions and stronger staining differences compared to natural images, and target morphologies often suffer from problems such as blurred boundaries, local adhesion, internal holes, and strong background interference. While traditional U-Net-like methods can achieve good foreground localization results in the overall region, they still have shortcomings in areas such as shallow texture noise suppression, cross-layer context memory utilization, fine-grained boundary restoration, and separation of adhered targets. Therefore, there is an urgent need to provide a new pathological image segmentation method that can further improve shallow noise suppression, cross-layer feature memory modeling, boundary detail restoration, and adaptive correction of complex structural errors in the post-processing stage while maintaining overall region recognition capabilities. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of insufficient shallow feature noise suppression, inadequate utilization of encoded features, weak fine-grained boundary recovery ability, and poor post-processing adaptability in existing pathological image segmentation methods, and to provide a pathological image segmentation method based on fractional memory enhancement and differential injection.

[0005] A further objective of this invention is to provide a complete pathological image segmentation process that enables the coordinated work of feature modulation within the network and structural refinement after network output, thereby improving the ability to preserve boundaries, separate adhered targets, repair holes, and suppress false detections in pathological image segmentation tasks.

[0006] To achieve the above objectives, this invention provides a pathological image segmentation method based on fractional-order memory enhancement and differential injection, comprising the following steps: S1. Obtain the pathological image to be segmented, and input the pathological image into the fractional-order memory-enhanced encoding-decoding segmentation network after standardization preprocessing. S2. Perform shared fractional-order pre-denoising processing on the shallow features obtained in the encoding stage, generate a smooth response through a learnable fractional-order smoothing kernel, and combine edge gating to suppress noise in non-boundary regions to obtain cleaned shallow features. S3. In the decoding stage, based on the corresponding decoding query features, fractional-order memory-weighted fusion is performed on multiple coding layer features to obtain memory-enhanced jump features; S4. Construct a shared fractional-order differential response based on shallow features, and inject the differential response into the basic jump features in the form of residuals in at least one decoding stage using boundary gating, separation gating and injection strength control mechanisms to enhance the boundary repair and adjacent target separation capabilities. S5. Utilize the fractional-order memory-enhanced encoding-decoding segmentation network to output a foreground probability map or equivalent segmentation prediction result for the pathological image. S6. Perform candidate mask reconstruction on the foreground probability map, and construct multiple counterfactual candidate results for the candidate connected components, including at least the candidate to be retained, the candidate to be repaired, the candidate to be split, and the candidate to be suppressed. S7. Based on the joint energy evaluation function, compare and select the multiple counterfactual candidate results, and output the final pathological image segmentation mask.

[0007] Preferably, in step S2, the shared fractional-order pre-denoising process is applied at least to the first layer of shallow coding features and can be further applied to the second layer of shallow coding features; the order of the fractional-order smoothing kernel is a learnable parameter, and the edge gating is generated by local contrast or equivalent local edge strength information to reduce the smoothing intensity of strong boundary regions.

[0008] Preferably, in step S3, the fractional-order memory weighted fusion includes: aligning the encoded features involved in the fusion by channel alignment and spatial scale alignment, constructing basic memory weights according to the fractional-order memory decay law, obtaining the final fusion weights by combining the dynamic bias driven by the decoded query features, and performing channel modulation and convolution fusion on the fusion result.

[0009] Preferably, in step S4, the shared fractional differential response is obtained by performing fractional differential operations on shallow features in at least two directions and is shared by multiple decoding stages; each decoding stage generates boundary gating, separation gating and injection intensity according to the current basic features, query features and prior context, so as to control the injection position and amplitude of differential residuals.

[0010] Preferably, in step S6, the candidate mask reconstruction includes at least one of dual-threshold hysteresis reconstruction, small target removal, hole filling, and boundary band correction.

[0011] Preferably, in step S7, the joint energy evaluation function considers at least two of the following evaluation items: confidence consistency energy, boundary alignment energy, topological consistency energy, separation rationality energy, and shape regularity energy.

[0012] The present invention also provides a pathological image segmentation apparatus, the apparatus comprising at least one processor and at least one memory, the memory storing program instructions, which, when executed on the processor, implement the above-described pathological image segmentation method.

[0013] The present invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the above-described pathological image segmentation method.

[0014] Beneficial effects: Compared with the prior art, the present invention has at least the following beneficial effects: 1. By sharing a fractional-order pre-denoising mechanism, the background texture interference and the risk of coloring noise propagation are reduced while the boundary information of shallow features are preserved, providing more stable detailed input for subsequent decoding stages.

[0015] 2. By using a fractional-order memory skip connection mechanism, features from multiple coding layers are adaptively fused according to fractional-order memory rules and decoding query conditions, which is more conducive to recovering complex contours and multi-scale structures than traditional single-layer skip connections.

[0016] 3. By sharing fractional-order differential response and stage differential injection mechanism, the network's response to fine-grained edges, narrow gaps and adjacent target boundaries is enhanced, thereby improving the separation effect of adhered targets and reducing undersegmentation.

[0017] 4. By using counterfactual candidate generation and joint energy evaluation mechanisms, post-processing no longer relies on a single fixed rule, but can adaptively select between preservation, repair, splitting and suppression for different connected domains, thereby improving the ability to correct false detections, missed detections, holes and adhesion errors.

[0018] 5. This invention integrates the design of in-network feature modulation and out-of-network result refinement to form a complete technical link, which is suitable for segmentation tasks of glands, cell nuclei and other tissue targets in pathological images. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the pathological image segmentation method of the present invention; Figure 2 This is a schematic diagram of the shared fractional-order pre-denoising module of the present invention; Figure 3 This is a schematic diagram of the fractional-order memory jump connection module of the present invention; Figure 4 This is a schematic diagram of the shared fractional-order differential response and staged differential injection module of the present invention; Figure 5 This is a schematic diagram of the counterfactual post-processing flow of the present invention; Figure 6 This is a schematic diagram showing the visualization results of the present invention in the task of pathological image segmentation. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Unless otherwise specified, the following embodiments and technical features described herein can be combined with each other.

[0021] Example 1: Overall Process. This example provides a pathological image segmentation method based on fractional-order memory enhancement and differential injection. The overall process is as follows: Figure 1 As shown.

[0022] First, the pathological image to be segmented is acquired, and preprocessing operations such as size unification, pixel normalization, and label alignment are performed on the image. The processed image is then input into a fractional-order memory-enhanced encoder-decoder segmentation network. The network includes an encoder, a bottleneck layer, a decoder, a memory-enhanced skip connection branch, a shared fractional-order pre-denoising branch, and a shared fractional-order difference branch.

[0023] The encoding stage extracts texture, edge, and high-level semantic features from the pathological image layer by layer; the decoding stage restores the spatial resolution layer by layer and fuses the memory-enhanced skip features from the encoding stage. The network outputs foreground segmentation logits, which are then transformed by a sigmoid algorithm to obtain the foreground probability map of the pathological image.

[0024] Subsequently, counterfactual post-processing is performed on the probability graph. In the post-processing stage, the initial candidate mask is first reconstructed based on the probability distribution, then candidate masks for preservation, repair, splitting, and suppression are constructed for each connected component, and finally, a joint energy evaluation function is used for selection, outputting the final segmentation mask.

[0025] Through the above process, the present invention realizes a complete processing link from the original pathological image to the structural refinement and segmentation result.

[0026] Example 2: Shared fractional-order pre-denoising module. In pathological images, the first and second layer shallow coding features typically contain a lot of local texture details, but are also most prone to introducing background noise, staining fluctuations, and local artifacts. To reduce noise propagation, this invention sets up a shared fractional-order pre-denoising module on the shallow features.

[0027] For a given shallow feature, a fractional-order smoothing kernel is first constructed based on learnable fractional-order parameters, and split smoothing is performed on the feature along at least two directions to obtain a smoothed response. Then, an edge gating map is constructed based on the local mean and local contrast of the feature, ensuring that areas with strong edges are less affected by smoothing, while flat areas allow for more adequate noise suppression. Finally, the smoothing residual is modulated using learnable smoothing intensity coefficients to obtain the cleaned shallow feature. ; in, Indicates shallow features; Indicates a smooth response; Represents an edge-gated graph; This represents the learnable strength parameter; This indicates element-wise multiplication.

[0028] This structure allows shallow features to be smoothed in non-boundary regions while preserving the original structure as much as possible in boundary regions, thus achieving both noise suppression and edge preservation.

[0029] In one preferred embodiment, the shared fractional-order pre-denoising module operates on the first layer of shallow coding features; in another preferred embodiment, the shared fractional-order pre-denoising module operates on both the first and second layers of shallow coding features. The specific activation level can be set according to the pathological image type, staining stability, and target scale.

[0030] Example 3: Fractional-order memory skip connection module. Traditional skip connections typically only pass the encoded features of the same scale directly to the decoder. Although this method is simple, it is difficult to express the contextual memory relationship between different scales. Therefore, this invention sets up a fractional-order memory skip connection module.

[0031] For a given decoding stage, multiple coding layer features related to that stage are selected, and channel projection and spatial alignment are performed on them respectively to form a set of features to be fused. Then, basic memory weights are constructed according to the fractional-order memory principle. Preferably, the basic weights can satisfy a power-law decay with the temporal or hierarchical position of the features, so that features closer to the current decoding stage and earlier features participate in the fusion together according to the fractional-order memory method.

[0032] Furthermore, this invention introduces a dynamic bias term driven by the current decoding query features to modify the aforementioned basic memory weights, thereby obtaining the final fusion weights related to the current decoding context. Thus, multiple coding layer features are no longer fused in a fixed ratio, but rather adaptively change with the current decoding state.

[0033] After obtaining the weighted fusion features, channel gating and convolutional fusion are used to generate the final jump features. The final jump features and the decoding upsampling features are input together into the decoding fusion module to generate the output of the current stage.

[0034] This structure allows the decoding stage to comprehensively utilize information from multiple coding layers, giving it an advantage over single skip connections in restoring complex edges, hollow structures, and multi-scale tissue contours in pathological images.

[0035] Example 4: Shared fractional-order differential branch and staged differential injection module. In order to further enhance the network's response to boundaries, narrow gaps and adhesion regions, this invention sets up a shared fractional-order differential branch and injects its output into at least one decoding stage.

[0036] 1. Shared fractional-order difference response: This invention constructs a shared fractional-order difference response based on shallow features. Specifically, fractional-order difference operations are performed on the shallow features along the horizontal and vertical directions to obtain a directional difference response. Preferably, the fractional-order difference can be implemented based on truncation coefficients of the Grunwald-Letnikov form to balance directional detail representation and computational stability.

[0037] A detailed prior representation can be constructed from the differential responses and their amplitude information in at least two directions. This representation includes both boundary variation information and potential separation cues. This detailed prior, as a shared differential response, is reused across multiple decoding stages, rather than being constructed independently in each stage, thereby reducing redundant computation and improving cross-stage consistency.

[0038] Phased Differential Injection: In a certain decoding stage, basic jump features and current decoding query features are first obtained. Then, candidate residuals for repair are generated using the basic jump features, decoding query features, and shared differential responses. To avoid noise amplification caused by indiscriminate enhancement, this invention further generates boundary gating maps and separation gating maps: the boundary gating map is used to emphasize the role of candidate residuals in the boundary region, and the separation gating map is used to emphasize the role of candidate residuals at potential adhesion boundaries.

[0039] Subsequently, an injection strength coefficient is generated based on the decoded query features to control the overall injection amplitude of the differential residuals in this stage. The final enhancement feature can be represented as: ; in, Indicates the injection intensity coefficient; Indicates the basic jump characteristics; Indicates candidates for repairing residuals; Represents a boundary-gated graph; This represents a delimited gating graph.

[0040] In a preferred embodiment, staged differential injection is applied to at least one or more decoding stages near the output to enhance boundary recovery and separation of contiguous targets. In another preferred embodiment, differential injection is not enabled in deeper decoding stages to reduce interference with the coarse semantic reconstruction stage.

[0041] Example 5: Counterfactual candidate post-processing module. This invention sets up a counterfactual candidate post-processing module after the network output to perform structural-level adaptive correction for each connected component.

[0042] 1. Initial candidate mask reconstruction: For the foreground probability map output by the network, an initial candidate mask is first generated using a dual-threshold hysteresis reconstruction strategy. Specifically, a high-threshold region is used as a reliable seed, and a low-threshold region is used as a scalable support domain. Only regions connected to the reliable seed are retained, thereby minimizing noise introduction while preserving the integrity of the target as much as possible.

[0043] Afterwards, further processing such as morphological opening, small target removal, hole filling, and boundary band refinement can be performed to obtain a more stable initial connected region.

[0044] 2. Counterfactual candidate construction: For each connected component in the initial candidate mask, the present invention constructs at least the following candidate results: (1) Retain candidates: Keep the original connected components unchanged; (2) Suppress candidates: Remove the connected component for small regions with low confidence or suspected false positives; (3) Repair candidates: Relax the low threshold, perform closing operation or connectivity constraint in a local area to restore the boundary or internal missing regions that may have been missed; (4) Splitting candidates: For connected domains that may consist of multiple cohesive targets, splitting is performed using distance peaks, local valleys and watershed cues.

[0045] The above candidate results can be regarded as multiple counterfactual results for the same connected domain, that is, an explicit candidate set of "what structural results will be obtained if different correction strategies are adopted".

[0046] 3. Selection of Joint Energy Assessment: To select the most reasonable result from multiple candidate results, this invention defines a joint energy evaluation function. The evaluation function includes at least two of the following components: (1) Confidence Consistency Energy: measures the probability mean within a candidate region and the probability rationality of adding or removing candidate regions; (2) Boundary alignment energy: measures the degree of matching between the candidate boundary and the image gradient and the probability stability of the boundary band; (3) Topological consistency energy: measures the consistency between the number of connected components and the proportion of holes in the candidate results and the expected structure; (4) Separation rationality energy: measures the sufficiency and stability of candidate results in separating potential adhesion targets; (5) Shape regularity energy: measures the degree of convexity loss, contour irregularity and abnormal growth ratio of candidate results.

[0047] By weighted summing of each energy component, the total energy value of each candidate result can be obtained, and the candidate result with the smallest total energy is selected as the final output result of the connected domain.

[0048] 4. Regionalization and Refinement: In a preferred embodiment, the present invention can also construct local ROIs for unstable boundary regions, fragmented regions, or potential splitting regions, and repeatedly perform candidate construction and energy selection processes within these ROIs to improve local refinement efficiency and reduce excessive modification of stable regions of the whole map.

[0049] Example 6: Training and Inference Configuration. In a preferred embodiment, the network of this invention uses a combination of binary cross-entropy loss and Dice loss to optimize the main output, and sets up deep supervision branches in multiple intermediate decoding stages. For pathological image segmentation tasks with high boundary quality requirements, boundary-assisted loss can also be introduced to enhance the learning ability of contour structures.

[0050] The total loss function can be expressed as: ; in: Indicates the primary segmentation loss; This represents the auxiliary loss of the deep supervision branch; This indicates a deep supervised branch index; Indicates the auxiliary loss weight; This represents the boundary auxiliary loss weight.

[0051] In a preferred embodiment, shallow shared fractional-order pre-denoising, fractional-order memory jump connections, shared fractional-order differential branches, and stage differential injection can be enabled or disabled according to the characteristics of the dataset; preferably, they are implemented as modular switches so as to adjust the network structure for different pathological image types without changing the overall technical concept of the present invention.

[0052] In the inference phase, in addition to single-view prediction, a test-time enhancement strategy can be introduced. Preferably, the original image, horizontally flipped image, vertically flipped image, and double-flipped image are predicted separately, and a probability fusion method based on view confidence is used to obtain the final predicted probability map. In another preferred embodiment, post-processing can be performed on the prediction results of each view first, and then a voting strategy can be used to obtain the final mask.

[0053] Example 7: Experiment and Effect Verification. To verify the effectiveness of the method of the present invention, comparative experiments were conducted using three pathological image segmentation datasets: MoNuSeg, GlaS, and BCSS. The method of the present invention was also compared with networks such as UNet, UNet++, DCA-UNet, CM-UNeXt, I2U-Net, UNeXt, DMSA-UNet, MSFLUNet, and EGNet. Evaluation metrics included Dice, IoU, Precision (Pre), Recall, Specificity (Spec), and Accuracy (Acc). In this example, the comparison methods were performed under the same data partitioning, the same evaluation metrics, and consistent or similar experimental conditions. Tables 1 to 3 present the quantitative results on the three datasets. Some qualitative visualization results of the method of the present invention are shown in the accompanying figures. Figure 6 Provided.

[0054] 1. Comparison results on the MoNuSeg dataset As shown in Table 1, the method of this invention achieves Dice, IoU, Pre, Spec, and Acc metrics of 81.62%, 69.25%, 79.44%, 94.51%, and 92.66% respectively on the MoNuSeg dataset, obtaining the best results among the comparative methods listed in this embodiment. The Recall is 84.75%, slightly lower than DMSA-UNet's 86.10%, but still maintains a high level overall. Compared with the suboptimal results, the method of this invention improves Dice and IoU by 0.72 and 0.75 percentage points respectively; compared with the basic UNet, they improve by 4.42 and 4.95 percentage points respectively, indicating that the method of this invention has significant advantages in maintaining cell nuclear boundaries and structural integrity.

[0055] Table 1 Performance comparison of different network segmentation accuracies on the MoNuSeg dataset 2. Comparison results on the GlaS dataset As shown in Table 2, the proposed method achieves Dice, IoU, Pre, and Acc metrics of 88.64%, 82.08%, 89.32%, and 90.93% respectively on the GlaS dataset, obtaining the best results among the comparative methods listed in this embodiment. The proposed method improves Dice, IoU, Pre, and Acc by 0.72 percentage points, 1.10 percentage points, 0.20 percentage points, and 0.75 percentage points respectively. These results demonstrate that the proposed method balances segmentation accuracy, boundary restoration, and false detection suppression in gland segmentation tasks, exhibiting superior overall performance.

[0056] Table 2 Performance comparison of different network segmentation accuracies on the GlaS dataset. 3. Comparison results on the BCSS dataset As shown in Table 3, in more complex BCSS histopathological scenarios, the proposed method achieves Dice, IoU, Pre, and Acc of 87.72%, 81.90%, 93.04%, and 87.65%, respectively, which are the best results among the comparative methods listed in this embodiment. Recall reaches 86.50%, only 0.29 percentage points different from the best DMSA-UNet (86.79%). While the Spec is 60.87%, it is not the best, but considering its higher Dice, IoU, Pre, and Acc comprehensive indicators, it can be seen that the proposed method maintains strong foreground detection capability while exhibiting superior overall segmentation performance, indicating its good robustness in target region identification under complex tissue backgrounds.

[0057] Table 3 Performance comparison of different network segmentation accuracies on the BCSS dataset The comparative experimental results of the three datasets above show that the method of this invention exhibits strong generalization ability and stability in different types of pathological image segmentation tasks. Figure 6 The visualization results show that the method of the present invention has good technical effects in terms of target boundary preservation, separation of adhered targets, hole repair and false detection suppression.

[0058] This invention provides a pathological image segmentation method based on fractional-order memory enhancement and differential injection. Addressing technical challenges in pathological image segmentation such as shallow noise, blurred boundaries, target adhesion, and poor post-processing adaptability, this method constructs a unified technical solution comprising shared fractional-order pre-denoising, fractional-order memory skip connections, shared fractional-order differential response, staged differential injection, and counterfactual post-processing. Through these technical means, this invention can improve boundary quality, separation capability, and structural rationality while maintaining overall region recognition ability, making it applicable to various pathological image segmentation scenarios.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pathological image segmentation method based on fractional-order memory enhancement and differential injection, characterized in that, Includes the following steps: S1. Input the pathological image into the fractional-order memory-enhanced encoding-decoding segmentation network, perform shared fractional-order pre-denoising processing on the shallow representation of the pathological image, and use edge gating to suppress the effect of fractional-order smoothing in the edge region to obtain the cleaned shallow features. S2. Shallow memory features and shared fractional differential responses are synchronously formed from the cleaned shallow features; S3. In the decoding stage, the fractional basis weights that decay with memory position are corrected by the stage query features to obtain the fusion weights, and the encoding side features and the shallow memory features are weighted and fused according to the fusion weights. S4. Under the joint constraints of boundary gating, separation gating and query-driven injection strength, the aligned shared fractional differential response is injected into the basic features in a residual manner, and the target foreground response map is formed based on the decoded features after weighted fusion and residual injection correction. S5. Reconstruct candidate masks based on the target foreground response map, and construct counterfactual candidate results for candidate connected regions, including retained candidates and one or more structurally modified candidates, wherein the structurally modified candidates include at least one of repair candidates, split candidates and suppression candidates. S6. Based on the joint energy, compare and select the counterfactual candidate results, and output the pathological image segmentation mask.

2. The pathological image segmentation method according to claim 1, characterized in that, The shared fractional-order pre-denoising process includes: mapping shallow features to a preset channel dimension, generating a fractional-order smoothing kernel based on learnable fractional-order parameters; smoothing the shallow features using the fractional-order smoothing kernel to obtain fractional-order smoothed features; constructing an edge response based on the difference between the shallow features and their local mean, and generating an edge gating from the edge response; controlling the fusion strength between the fractional-order smoothed features and the original shallow features based on the edge gating, so that the edge region retains a higher original response, and the non-edge region is enhanced with denoising and smoothing.

3. The pathological image segmentation method according to claim 1, characterized in that, The fractional-order memory weighted fusion includes: constructing negative fractional power decay base weights for each memory position participating in the fusion according to the position index; mapping the base weights to the logarithmic domain, superimposing them with the slot bias set for each memory position and the dynamic correction amount generated by the query features of the current decoding stage, and normalizing them to obtain the fusion weight corresponding to the current position; performing weighted summation on the multi-layer coding features and shallow memory features according to the fusion weights, and forming the memory-enhanced skip connection feature through channel mapping; wherein, the fusion weights simultaneously reflect the proximity of the memory position and the semantic requirements of the current decoding stage.

4. The pathological image segmentation method according to claim 1, characterized in that, The construction of the shared fractional-order difference response includes: generating fractional-order difference coefficients based on the cleaned shallow features; performing fractional-order difference operations on the cleaned shallow features along at least two directions using the fractional-order difference coefficients to obtain directional difference responses; aggregating the amplitudes of the difference responses in different directions to obtain a shared fractional-order difference response that simultaneously characterizes the edge intensity and the magnitude of detail changes; wherein the fractional-order difference coefficients are generated recursively according to the series expansion of fractional-order differences.

5. The pathological image segmentation method according to claim 1, characterized in that, The staged differential injection includes: generating a differential enhancement amount based on the basic features, query features, and shared fractional differential responses; generating a control signal including at least boundary gating and separation gating based on the basic features, query features, and prior context, and generating an injection strength from the query features; and superimposing the differential enhancement amount modulated by the control signals and injection strength onto the basic features in a residual manner.

6. The pathological image segmentation method according to claim 1, characterized in that, The construction of the counterfactual candidate results includes: reconstructing an initial candidate mask based on the target foreground response map, and constructing at least one candidate result for structural correction for the candidate region; the candidate result for structural correction includes one or more of repair candidates, split candidates and suppression candidates, wherein the repair candidate is used to complete the candidate region, the split candidate is used to separate the adhered region, and the suppression candidate is used to remove the low confidence region.

7. The pathological image segmentation method according to claim 1, characterized in that, The joint energy is composed of at least two of the following: confidence consistency within the region, consistency of candidate boundaries, rationality of candidate topology, rationality of target separation, and rationality of candidate shape. The retained candidates are used as the baseline candidates, and the candidate results used for structural correction are compared. When the joint energy improvement conditions are met, the corresponding candidates are used to form the final pathological image segmentation mask.

Images