Weakly Supervised Whole Slide Image Screening Without Pixel-Level Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high clinical and laboratory workload in histopathology laboratories, coupled with a shortage of pathologists and MLTs, leads to increased human error and resource diversion due to the impracticality of manual pixel-level annotation for training deep learning models, hindering efficient cancer diagnosis and triage.
Innovation Solution
A weakly supervised machine-learning technique using ensembled networks with optimized hyperparameters processes whole slide images (WSIs) to classify and prioritize cancer cases, utilizing averaged cellular features without extensive pixel-level annotation, and generates tumor probability heatmaps for visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual pixel-level annotation is used for training deep learning models, then model accuracy is improved, but time consumption and human resource requirements increase significantly
Solution Approach 1:
The patent extracts only the essential information needed for training by using weakly supervised learning with slide-level labels instead of requiring pixel-level annotations. This extracts the core diagnostic signal while discarding the time-consuming annotation process, resolving the contradiction between accuracy and time consumption.
Solution Approach 2:
The patent introduces an intermediary attention mechanism that automatically identifies and focuses on relevant regions within WSIs without manual annotation. This intermediary component bridges the gap between coarse slide-level labels and fine-grained diagnostic features, achieving accurate model training without pixel-level annotation effort.
2Measurement precision
If manual pixel-level annotation is performed, then training data quality is improved, but labor resources and cost increase
Solution Approach 1:
The system performs self-service by automatically generating training signals from slide-level diagnoses without requiring manual pixel-level annotation. The attention mechanism and loss function work together to self-identify relevant regions and self-train the model, eliminating the need for extensive manual labor while maintaining training data quality.
3Measurement precision
If pathologists review all cases manually, then diagnostic accuracy is maintained, but productivity decreases due to overwhelming workload
Solution Approach 1:
The patent segments the diagnostic workflow into two parts: an AI model handles initial screening and prioritization of WSIs, while pathologists focus only on reviewing high-priority cases flagged by the model. This segmentation maintains diagnostic accuracy for critical cases while dramatically improving overall productivity by reducing the volume of cases requiring manual review.
4Measurement precision
If extensive pixel-level annotation is required, then model training effectiveness is improved, but ease of manufacture and deployment deteriorates
Solution Approach 1:
The patent changes the annotation parameter from fine-grained pixel-level labels to coarse-grained slide-level labels. This parameter change simplifies the data collection and model deployment process while maintaining training effectiveness through the attention mechanism that automatically recovers fine-grained information from coarse labels.
Data Source
AI summary
In generating training and testing datasets for machine learning, manual annotation of a large volume of data is impractical and expensive. A machine-learning model for detecting carcinoma (CA) from a whole slide image (WSI) processes average cellular features of cells identified on the WSI. Each cellular feature is a descriptive statistic of the cells, advantageously allowing the training and testing datasets to be constructed without a costly annotation process of pixelwise labelling each cell on a WSI training sample. Apart from predicting a CA case or a non-CA case for the WSI, the machine-learning model is also usable to: identify a suspicious CA case for priority assessment if a non-CA case is predicted for the WSI; generate a tumor probability heatmap of the WSI for visualizing potential CA regions on the WSI to assist pathological assessment; and assess quality control of a triage system before implementation in clinical setting.