Biomedical Image Survival Scoring Through End-to-End Histology Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current survival models for histopathology-based cancer prognosis struggle to provide stratified patient groups driven by histologic morphologies and are limited by two-stage training frameworks that decouple image information, especially in rare cancers like intrahepatic cholangiocarcinoma (ICC).
Innovation Solution
EPIC-Survival employs an end-to-end deep learning approach integrating tile encoding and aggregation with stratification boosting to directly produce survival risk scores from whole slide images (WSIs), enhancing the model's ability to identify specific histologic features and form well-defined risk groups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If two-stage training frameworks are used to process histopathology images, then the processing pipeline is modular and easier to implement, but the model cannot effectively capture integrated image information and form well-defined risk groups
Solution Approach 1:
The patent merges the encoding and aggregation stages into a unified end-to-end training framework. The model processes histopathology images through integrated tile encoding and aggregation operations, allowing information to flow continuously from input to output without decoupling. This unified architecture enables the model to capture comprehensive image information and form well-defined risk groups, directly resolving the limitation of two-stage frameworks while maintaining implementation feasibility through a single training process.
2Device complexity
If two-stage training frameworks are used, then the model architecture is simpler to construct, but the model loses ability to systematically map tissue morphology to patient outcomes
Solution Approach 1:
The patent implements continuous information flow through end-to-end training, where tile encoding and aggregation operations proceed seamlessly from input images to output predictions. The unified model architecture maintains continuous mapping from tissue morphology features to patient outcomes, eliminating the information loss that occurs in decoupled two-stage frameworks. This continuity enables systematic mapping while the model architecture remains constructible through standard deep learning practices.
3Adaptability or versatility
If current survival models are used for rare cancers, then general survival analysis can be performed, but the models cannot provide stratified patient groups driven by specific histologic morphologies
Solution Approach 1:
The patent applies local quality by enabling the model to identify and weight specific histologic morphology features relevant to rare cancers like intrahepatic cholangiocarcinoma. The end-to-end training framework allows different regions and features of the histopathology images to contribute differently to the prediction, with the model learning to focus on morphology patterns characteristic of rare cancer types. This produces well-defined risk groups stratified by specific histologic features while maintaining applicability to rare cancers through the generalizable end-to-end architecture.
Data Source
AI summary
Presented herein are systems and methods for determining scores from biomedical images. A computing system may identify a plurality of tiles in a first biomedical image derived from a sample of a subject. Each tile may correspond to features of the sample. The computing system may apply the plurality of tiles to a machine learning (ML) model. The ML model may include: an encoder to generate a plurality of feature vectors based on the plurality of tiles; a clusterer to select a subset from the plurality of feature vectors; and an aggregator to determine a first score indicative of a time to an event for the subject resulting from the features of the sample. The model may be trained in accordance with a loss derived from second scores determined for second biomedical images. The computing system may store an association between the score and the first biomedical image.