Histological Image Analysis via Multi-Resolution Tile Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for analyzing histological images, particularly for predicting patient outcomes in colorectal cancer, are inconsistent and lack robustness for clinical application.
Innovation Solution
A computer-implemented system that generates multiple tiles from histological images with varying areas and resolutions, processed by machine-learning networks to determine classifiers, which are then combined to produce an overall classifier for improved prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single classifier is used for histological image analysis, then the system is simple, but the prediction accuracy and consistency are insufficient for clinical application
Solution Approach 1:
The patent combines multiple classifiers (first classifier from low-resolution tiles and second classifier from high-resolution tiles) into an overall classifier through a classifier combiner. This merging approach integrates the strengths of different classification models to achieve higher prediction accuracy and consistency, resolving the contradiction between reliability and complexity by showing that the improved accuracy justifies the additional computational complexity.
2Measurement precision
If high-resolution tiles are used for detailed analysis, then measurement precision improves, but processing time and computational resources increase
Solution Approach 1:
The patent segments the histological image into multiple tiles at different resolutions. Low-resolution tiles are processed first to capture broad patterns and generate a first classifier, while high-resolution tiles are processed separately to capture detailed features for a second classifier. This segmentation allows parallel processing and optimizes the trade-off between precision and processing time by assigning different computational resources to different resolution levels.
Solution Approach 2:
The patent introduces a resolution dimension by processing tiles at multiple resolution levels (low-resolution and high-resolution). This multi-dimensional approach allows the system to simultaneously achieve efficient processing (through low-resolution overview) and high precision (through high-resolution detail analysis), resolving the time-precision trade-off by operating in multiple resolution dimensions rather than a single fixed resolution.
Data Source
AI summary
A computer implemented system for determining an overall-classifier for one or more source-histological-images. The system comprising: a first tile generator (204) configured to generate a plurality of first-tiles (206; 306) from the one or more source-histological-image (202; 302); and a second tile generator (205) configured to generate a plurality of second-tiles (207; 307) from the one or more source-histological-images (202; 302). The first-area of the first-tiles (206; 306) is larger than the second-area of the second-tiles (207; 307); and the second-resolution of the second-tiles (207; 307) is higher than the first-resolution of the first-tiles (206; 306). The system also includes a machine-learning network (211; 311) configured to process the plurality of first-tiles (206; 306) in order to determine a first-classifier (218; 318); a machine-learning network (215; 311) configured to process the plurality of second-tiles (207; 307) in order to determine a second-classifier (219; 319); and a classifier combiner configured to combine the first-classifier (218; 318) and the second-classifier (219; 319) to determine the overall-classifier (232; 332).