Whole-Slide Annotation Refinement for Weakly Supervised Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training machine-learning algorithms for whole-slide images in digital pathology is time-consuming and error-prone due to the high cost and complexity of obtaining accurate ground-truth labels, especially for diverse and rare tissue types, which current annotation methods fail to address efficiently.
Innovation Solution
A method for refining annotations of whole-slide images by employing patch-wise weak supervision and class activation maps to determine a refined annotation with a higher level of detail, facilitating faster and more accurate training of segmentation algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation of whole-slide images is performed to obtain ground-truth labels, then annotation accuracy is improved, but annotation time and cost increase significantly
Solution Approach 1:
The patent divides the whole-slide image into multiple patches and processes them independently. The segmentation algorithm processes patches in parallel, reducing the overall annotation time while maintaining accuracy through localized analysis of tissue structures in each patch.
Solution Approach 2:
The patent introduces a weakly-supervised segmentation algorithm as an intermediary tool that generates preliminary annotations to assist pathologists. This intermediary system pre-processes images and provides suggested segmentations, reducing the manual annotation burden while maintaining high accuracy through expert verification.
2Manufacturing precision
If detailed pixelwise ground-truth labels are obtained for all tissue types, then segmentation precision is improved, but annotation complexity and cost increase
Solution Approach 1:
The patent segments the annotation task into patch-level operations, where each patch is annotated independently with simplified labels. This reduces the overall complexity by breaking down the large-scale annotation problem into manageable units while maintaining segmentation precision through localized analysis.
Solution Approach 2:
The patent implements weakly-supervised learning where the algorithm learns from partially annotated data rather than requiring complete pixelwise labels for all tissue types. This partial annotation approach reduces complexity while still achieving high segmentation precision through iterative refinement and class activation mapping.
3Reliability
If comprehensive annotations covering all tissue types are obtained, then algorithm training accuracy is improved, but resource requirements and cost increase
Solution Approach 1:
The patent performs preliminary weakly-supervised segmentation to generate class activation maps and preliminary annotations before final algorithm training. This preliminary action identifies important tissue regions and creates initial training data, reducing the resources needed for comprehensive manual annotation while maintaining training accuracy.
Solution Approach 2:
The patent uses the weakly-supervised segmentation algorithm to generate synthetic training annotations that serve as copies of ground-truth labels. These copied annotations cover diverse tissue types and provide sufficient training data without requiring expensive manual annotation for every sample, thereby reducing resource requirements while maintaining training reliability.
Data Source
Figure 1~3
Figure 4~5
Figure 6~7
AI summary
Various disclosed examples pertain to digital pathology, specifically to training of a segmentation algorithm for segmenting whole-slide images (111) depicting tissue of multiple types. An initial annotation (120) of a whole-slide image is refined to yield a refined annotation based on which parameters of the segmentation algorithm can be set. Techniques of patch-wise weak supervision can be employed for such refinement.