CN-122025038-A - Automatic labeling system and method for pathological image

CN122025038ACN 122025038 ACN122025038 ACN 122025038ACN-122025038-A

Abstract

The invention discloses an automatic labeling system and method for pathological images, which relate to the technical field of image labeling, wherein a strategy optimization module inputs a comprehensive characteristic data set and disturbance factors into a labeling prediction model, takes characteristic change indexes and labeling uncertainty indexes as structural constraints, obtains space-time distributed labeling prediction data and a full-graph key analysis area list through multiple iterations, and a dynamic labeling module synthesizes the full-graph key analysis area distribution, the characteristic change indexes and the labeling uncertainty indexes of each unit to form a regional suitability map, acquires image use region distribution data, and dynamically adjusts analysis strength and isolation strategies of a labeling process according to the regional suitability map and the use region distribution data. The marking system realizes cooperative improvement of marking precision and efficiency of pathological images by organically combining multi-source feature fusion, constraint iterative optimization and dynamic strategy adjustment, and provides a high-efficiency solution for automated analysis of pathological images.

Inventors

LIU YUANZHU
LI MAOLIN
XIAO LEYI
QIU NA
ZHANG TENGYE
YANG QINYING
Lv Jihao
PAN HUIYUN
WANG XUEWEN
GAO MING

Assignees

青岛言鼎盛医疗器械科技有限公司
青岛言鼎生物医疗科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260115

Claims (10)

1. The automatic labeling system for the pathological image is characterized by comprising a data set generation module, a strategy optimization module and a dynamic labeling module: The data set generation module is used for uniformly dividing a pathological image to be marked into a plurality of space grid units, collecting multi-source characteristic data to form a multi-source characteristic data set, calculating disturbance factors including characteristic change indexes and marking uncertainty indexes, weighting and fusing the multi-source characteristic data according to the disturbance factors, and generating a comprehensive characteristic data set; The strategy optimization module inputs the comprehensive characteristic data set and the disturbance factor into a labeling prediction model, takes the characteristic change index and the labeling uncertainty index as structural constraints, and obtains labeling prediction data of space-time distribution and a full-graph key analysis region list through multiple iterations; and the dynamic labeling module is used for synthesizing the overall map key analysis region distribution, the characteristic change index and the labeling uncertainty index of each unit to form a region suitability map, acquiring image use region distribution data, and dynamically adjusting the analysis intensity and the isolation strategy of the labeling process according to the region suitability map and the use region distribution data.
2. The automatic labeling system for pathological images according to claim 1, wherein the strategy optimization module inputs the comprehensive feature data set and the disturbance factors into a labeling prediction model, strengthens boundary and detail fitting constraint on a high feature change region inside the labeling prediction model, introduces a multi-fusion and confidence attenuation mechanism on a high uncertainty region, and obtains labeling prediction data of space-time distribution and a key analysis region list through multiple iterations.
3. The system for automatically labeling pathological images according to claim 2, wherein the policy optimization module detects labeling confidence differences among adjacent grid cells in real time, and triggers the change of the strategy of the cross-region information flow path if the differences are greater than or equal to a first confidence difference threshold value: The strategy comprises the steps of prolonging a path of an intermediate abstraction layer for feature transfer, expanding a fusion space attenuation range, and arranging an embedded feature isolation unit at a logic boundary of an adjacent block, wherein the feature isolation unit vertically penetrates through a multi-scale feature layer in a data processing pipeline, and when a confidence coefficient difference triggers a threshold value, the feature isolation unit dynamically enhances a feature shielding coefficient.
4. The automatic labeling system for pathological images according to claim 2, wherein the strategy optimization module inputs the comprehensive feature data set and the disturbance factor into a labeling prediction model, takes a feature change index and a labeling uncertainty index as structural constraints, and obtains labeling prediction data of space-time distribution and a full-graph key analysis region list through multiple iterations: For a high-feature change region, the weight of a boundary and detail fitting constraint term is enhanced in a loss function, so that the network focuses on edge sharpness, shape continuity and texture details in the training and reasoning stage; for a high-annotation uncertainty region, introducing a multi-fusion mechanism, maintaining a plurality of hypothesis branches in parallel in a feature fusion stage, wherein each branch adopts different receptive fields or attention configurations, and outputting a plurality of candidate results for the annotation of the same position; Weighting and synthesizing each candidate result through a confidence coefficient attenuation mechanism, wherein the weight is determined by the fluctuation degree and uncertainty index of the time sequence characteristics of the region; The iteration process is carried out alternately by adopting multiple rounds of forward propagation and error reverse propagation, the difference between the labeling prediction result and the pseudo tag is evaluated after each round of iteration, and the network parameters and the constraint weight are updated; and finally, outputting labeling prediction data of the space-time distribution until the integral confidence coefficient and the space consistency of the prediction result reach a preset convergence condition.
5. The automated labeling system of pathology images of claim 3, wherein the policy optimization module enters a real-time detection phase to continuously monitor labeling confidence differences between neighboring grid cells: After each batch of reasoning is finished, calculating the absolute value of the difference between the labeling confidence scores of adjacent units, and if the absolute value is larger than or equal to a preset first confidence difference threshold value, judging that the risk of unbalance of the cross-region confidence is existed, triggering and changing the strategy of the cross-region information circulation path, wherein the method comprises the following steps: Extending the intermediate abstract layer path of feature delivery; And the attenuation range of the fusion space is enlarged.
6. The automatic labeling system for pathological images according to claim 2, wherein the dynamic labeling module dynamically adjusts the analysis intensity and isolation strategy of the labeling process: For the blocks with high suitability and located in the using area, the feature extraction depth and the attention focusing range are improved so as to capture the fine lesion structure; for low suitability or non-use areas, the operation complexity is reduced, a fusion strategy is adopted, the calculation force is saved, and overfitting is avoided.
7. The system for automatically labeling pathological images according to claim 6, wherein the dynamic labeling module dynamically updates constraint weights and fusion coefficients of the labeling prediction model according to the real-time confidence coefficient difference and disturbance factor change: If the characteristic change index of a certain region increases in the iterative process, enhancing the detail fitting constraint of the region and the intervention frequency of the characteristic isolation unit; if the uncertainty index continues to increase, a hypothetical branch and confidence re-estimation mechanism is introduced.
8. The automatic labeling system for pathological images according to claim 7, wherein the dynamic labeling module synthesizes the distribution of the full-image key analysis region and the characteristic change indexes and the labeling uncertainty indexes of each unit to form a region suitability map, acquires the distribution data of the image used region, and dynamically adjusts the analysis intensity and the isolation strategy of the labeling process according to the region suitability map and the distribution data of the used region: The foreground segmentation utilizes a semantic segmentation network or a threshold segmentation method to extract a pixel set containing a substantial tissue; For the blocks which simultaneously meet the high suitability level and are positioned in the using area, the deep feature extraction branches are started in the network by the blocks or the stacking number of the convolution layers is increased; For a tumor area marked as the highest suitability, connecting a detail enhancement branch outside a main network in parallel, and enabling a space attention unit to cover a neighborhood crossing a plurality of grid units; for blocks with low suitability or in non-use areas, the operation complexity is reduced; continuously tracking the confidence score of each block and the latest calculated characteristic change index and the labeling uncertainty index in the labeling execution process; when the fact that the characteristic change index of a certain area is increased in the iteration process is detected, the detail fitting constraint of the area is enhanced, the weight of the boundary and texture consistency penalty term is increased in the loss function, and the intervention frequency of the characteristic isolation unit is increased; if the texture of a certain region is complicated due to tissue folding, structural constraint is applied to the region in a plurality of subsequent iterations, and the activation times of the isolation units are improved; If the labeling uncertainty index of a certain area continuously increases, a plurality of parallel hypothesis branches are opened up for the area in the forward path of the network.
9. The automated labeling system of pathology images of claim 8, wherein the data set generation module calculates the perturbation factor comprising: The characteristic change index reflects the change speed of characteristic intensity along with analysis scale or resolution, the scale logarithmic relation between characteristic response in the characteristic profile model characterization neighbor pixel layer and the distance pixel center is constructed, the structure modulation factor is determined by combining the edge density and the gray gradient, the characteristic profile model is corrected, the characteristic intensity change rate among different scales is calculated through the characteristic profile model, and the characteristic change index is obtained; The marking uncertainty index reflects the fluctuation degree of the feature under the time sequence or multi-view condition, the feature profile model is utilized to generate time sequence features in a set time window, and the ratio of standard deviation to mean value of the time sequence features is calculated to obtain the marking uncertainty index.
10. An automatic labeling method for pathological images, which is realized by the labeling system as claimed in any one of claims 1-9, and is characterized in that the labeling method comprises the following steps: Uniformly dividing a pathological image to be marked into a plurality of space grid units, collecting multi-source characteristic data to form a multi-source characteristic data set, calculating disturbance factors including characteristic change indexes and marking uncertainty indexes, weighting and fusing the multi-source characteristic data according to the disturbance factors, and generating a comprehensive characteristic data set; the feature data set and the disturbance factors are input into a labeling prediction model, and feature change indexes and labeling uncertainty indexes are used as structural constraints, so that labeling prediction data of space-time distribution and a full-graph key analysis region list are obtained through multiple iterations; And (3) integrating the overall map key analysis region distribution, the characteristic change index and the labeling uncertainty index of each unit to form a region suitability map, acquiring image use region distribution data, and dynamically adjusting the analysis intensity and the isolation strategy of the labeling process according to the region suitability map and the use region distribution data.

Description

Automatic labeling system and method for pathological image Technical Field The invention relates to the technical field of image annotation, in particular to an automatic annotation system and method for pathological images. Background In the field of automatic labeling of pathological images, along with the rapid development of digital pathology and artificial intelligence technology, the requirements of clinical and scientific research on large-scale and high-precision image labeling are increasing, however, the traditional labeling method is mostly dependent on a simple global automation algorithm, and is difficult to cope with the problems of high heterogeneity of tissue structures in pathological images, obvious difference of local feature complexity, mixing of effective areas and invalid backgrounds and the like, so that labeling precision is unstable, and critical lesion areas are easy to miss-label or mislabel. Therefore, a technical scheme capable of finely characterizing image characteristics, dynamically optimizing labeling strategies and intelligently distributing resources is needed to meet the dual requirements of automatic labeling of pathological images on high precision and high efficiency. Disclosure of Invention The invention aims to provide an automatic labeling system and method for pathological images, which are used for solving the problem of the defects in the background technology. In order to achieve the purpose, the invention provides the following technical scheme that the automatic labeling system for the pathological image comprises a data set generation module, a strategy optimization module and a dynamic labeling module: The data set generation module is used for uniformly dividing a pathological image to be marked into a plurality of space grid units, collecting multi-source characteristic data to form a multi-source characteristic data set, calculating disturbance factors including characteristic change indexes and marking uncertainty indexes, weighting and fusing the multi-source characteristic data according to the disturbance factors, and generating a comprehensive characteristic data set; The strategy optimization module inputs the comprehensive characteristic data set and the disturbance factor into a labeling prediction model, takes the characteristic change index and the labeling uncertainty index as structural constraints, and obtains labeling prediction data of space-time distribution and a full-graph key analysis region list through multiple iterations; and the dynamic labeling module is used for synthesizing the overall map key analysis region distribution, the characteristic change index and the labeling uncertainty index of each unit to form a region suitability map, acquiring image use region distribution data, and dynamically adjusting the analysis intensity and the isolation strategy of the labeling process according to the region suitability map and the use region distribution data. In a preferred embodiment, the strategy optimization module inputs the comprehensive feature data set and the disturbance factor into the labeling prediction model, and in the labeling prediction model, the boundary and detail fitting constraint is strengthened for the high feature change region, a multi-fusion and confidence degree attenuation mechanism is introduced for the high uncertainty region, and the labeling prediction data and the key analysis region list of space-time distribution are obtained through multiple iterations. In a preferred embodiment, the strategy optimization module detects the labeling confidence difference between adjacent grid units in real time, and triggers the change of the strategy of the cross-region information flow path if the difference is greater than or equal to a first confidence difference threshold value: The strategy comprises the steps of prolonging a path of an intermediate abstraction layer for feature transfer, expanding a fusion space attenuation range, and arranging an embedded feature isolation unit at a logic boundary of an adjacent block, wherein the feature isolation unit vertically penetrates through a multi-scale feature layer in a data processing pipeline, and when a confidence coefficient difference triggers a threshold value, the feature isolation unit dynamically enhances a feature shielding coefficient. In a preferred embodiment, the strategy optimization module inputs the comprehensive feature data set and the disturbance factor into the annotation prediction model, takes the feature change index and the annotation uncertainty index as structural constraints, and obtains the annotation prediction data of the space-time distribution and the full-graph key analysis region list through multiple iterations: For a high-feature change region, the weight of a boundary and detail fitting constraint term is enhanced in a loss function, so that the network focuses on edge sharpness, shape continuity and texture details in the t