Regional Embeddings for Efficient Whole-Slide Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models for digital pathology face challenges due to the high spatial resolution and morphological variance of whole-slide images, leading to inefficient training and classification, especially when dealing with weakly annotated data.
Innovation Solution
A multiple-instance learning method that considers both selected patches and their regional context by generating patch and regional embeddings, using a pre-trained model for feature extraction, and combining these embeddings to form a global embedding for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If patch-level labeling is used for training deep learning models, then classification accuracy can be improved, but the labeling process becomes extremely time-consuming and may even be impossible for certain clinical scenarios
Solution Approach 1:
The whole-slide image is segmented into multiple patches, and the labeling task is segmented from the whole-slide level to the patch level. This allows the model to learn from individual patches while using efficient patch-level labeling instead of time-consuming whole-slide labeling, resolving the contradiction between accuracy and time consumption.
Solution Approach 2:
Instead of requiring complete whole-slide labeling, the method uses partial labeling at the patch level. By selecting and labeling only representative patches, the training process achieves good classification accuracy with significantly reduced labeling time and effort.
2Loss of information
If the entire whole-slide image is used as input for deep learning models, then complete information is available for analysis, but the high spatial resolution and multi-gigabyte size lead to inefficient training and computational challenges
Solution Approach 1:
The large whole-slide image is divided into smaller, manageable patches. This segmentation reduces the computational burden and memory requirements while maintaining the essential morphological information needed for accurate classification, thus improving training efficiency without significant information loss.
Solution Approach 2:
The method extracts relevant features and information from the whole-slide image by processing individual patches. This extraction approach captures the essential diagnostic information while eliminating the computational inefficiency of processing the entire high-resolution image at once.
3Productivity
If only individual patches are used for classification without considering regional context, then processing speed improves, but the ability to detect diseases and characterize tissues is reduced due to loss of spatial relationship information
Solution Approach 1:
The method merges information from multiple patches by grouping them into regions and then combining regional embeddings to form a global embedding. This combining process preserves spatial relationship information and improves disease detection accuracy while maintaining the processing efficiency benefits of patch-based approaches.
Solution Approach 2:
The approach adds a regional dimension to the patch-based classification. By organizing patches into regions and creating hierarchical embeddings (patch → region → global), the method captures spatial context without requiring processing of the entire high-resolution image, thus improving accuracy while maintaining speed.
Data Source
AI summary
Systems, methods, and computer programs disclosed herein relate to training a machine learning model and using the trained machine learning model to classify images, preferably medical images, using multiple-instance learning techniques. The machine learning model can be trained, and the trained machine learning model can be used for various purposes, in particular for the detection, identification and/or characterization of tumor types and/or gene mutations in tissues.