Weakly Supervised Histopathology Model Training via Segmented Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The performance of machine learning and deep learning models for histopathology is limited by the volume and quality of annotated examples required for training, making it expensive and time-consuming to ensure systems perform at a clinically relevant level and generalize across institutions.
Innovation Solution
The implementation of weakly supervised multi-label and multi-task learning methods allows for the training of deep learning models directly from pathologist diagnoses without extensive pixel-level annotations, reducing the need for extensive training labels by several orders of magnitude and enabling systems to predict specimen properties automatically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained with extensive pixel-level annotations to achieve high performance, then model accuracy improves, but the time and cost for data annotation increases significantly
Solution Approach 1:
The patent segments the annotation task into two levels: slide-level annotations (coarse) and tile-level annotations (fine). By first annotating at the slide level and then selectively annotating only the relevant tiles, the system reduces the total annotation workload while maintaining the ability to achieve high model accuracy through multi-scale processing.
Solution Approach 2:
The system performs preliminary slide-level annotations before tile-level annotations. This preliminary action identifies the regions of interest at the slide level, allowing subsequent tile-level annotations to focus only on relevant areas, thereby reducing the overall time and effort required for complete annotation.
2Adaptability or versatility
If machine learning models are trained with large volumes of annotated data to improve performance, then model generalization capability improves, but the cost and complexity of data preparation increases
Solution Approach 1:
The patent segments the data preparation process into slide-level and tile-level components. This segmentation allows the system to manage large volumes of data more effectively by processing and annotating data at different granularities, reducing the complexity of preparing and managing the complete dataset while maintaining model generalization capability.
Solution Approach 2:
The system uses a unified annotation platform that handles both slide-level and tile-level annotations, as well as both supervised and self-supervised learning tasks. This multi-functional approach simplifies data preparation complexity by using a single system for multiple purposes while enabling training on large diverse datasets for improved generalization.
3Reliability
If manual annotation is performed to ensure clinically relevant performance, then diagnostic accuracy improves, but the cost increases prohibitively
Solution Approach 1:
The patent segments the annotation process to perform coarse slide-level annotations followed by selective fine-grained tile-level annotations. This segmentation dramatically reduces the number of pixels that require manual annotation while maintaining the diagnostic accuracy needed for reliable pathology analysis, thereby reducing costs.
Solution Approach 2:
The system incorporates self-supervised learning components that automatically generate annotations or pre-process data without requiring manual intervention. This self-service capability reduces the overall cost by replacing expensive manual annotation for portions of the data while maintaining the reliability of the final diagnostic models.
4Measurement precision
If extensive training labels are used to train deep learning models, then model performance improves, but the time required for training increases
Solution Approach 1:
The patent segments the training process into stages: first training on slide-level annotations, then refining with tile-level annotations. This staged approach allows the model to learn high-level patterns quickly and then refine its understanding of detailed features, reducing total training time while maintaining high performance.
Solution Approach 2:
The system performs preliminary training with coarse slide-level labels before introducing fine tile-level labels. This preliminary action allows the model to converge to a reasonable performance level faster, and subsequent fine-tuning with detailed annotations refines the performance without requiring retraining from scratch, thereby reducing total training time.
Data Source
AI summary
Systems and methods are disclosed for receiving one or more electronic slide images associated with a tissue specimen, the tissue specimen being associated with a patient and/or medical case, partitioning a first slide image of the one or more electronic slide images into a plurality of tiles, detecting a plurality of tissue regions of the first slide image and/or plurality of tiles to generate a tissue mask, determining whether any of the plurality of tiles corresponds to non-tissue, removing any of the plurality of tiles that are determined to be non-tissue, determining a prediction, using a machine learning prediction model, for at least one label for the one or more electronic slide images, the machine learning prediction model having been generated by processing a plurality of training images, and outputting the prediction of the trained machine learning prediction model.