Tile-Based Multiple Instance Learning for Biomedical Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of large datasets and the computational complexity in training high-capacity models for biomedical image classification and localization in pathology has hindered the deployment of effective decision support systems in clinical settings.
Innovation Solution
A deep multiple instance learning (MIL) framework is applied to classify whole slide images (WSIs) using a convolutional neural network, where each WSI is treated as a bag of tiles, with only the slide-level class known, and the model learns from the top-ranked instance in each slide via a cross-entropy loss function, leveraging a large prostate needle biopsy dataset of 12,160 slides.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are trained on biomedical images with high resolution and large size, then classification accuracy and localization precision are improved, but computing resource consumption and training time increase significantly
Solution Approach 1:
The patent divides a whole slide image into multiple smaller tiles, processes each tile independently through the neural network, and then aggregates the results. This segmentation approach reduces the computational burden on each processing unit while maintaining the ability to achieve high classification accuracy through ensemble prediction across all tiles.
2Measurement precision
If deep learning models are trained on biomedical images with high resolution and large size, then classification accuracy and localization precision are improved, but training time increases significantly
Solution Approach 1:
By segmenting the large biomedical image into smaller tiles, the training process can be parallelized and each tile processed more quickly. The overall training time is reduced while the model still achieves high accuracy by learning from the aggregated predictions across multiple tiles.
Solution Approach 2:
The patent employs a two-stage training approach where the model is first trained on a subset of tiles or with simplified processing, then fine-tuned on the complete dataset. This partial action strategy allows the model to achieve good performance faster, with subsequent refinement improving accuracy without requiring exhaustive processing from the start.
3Ease of manufacture
If multiple instance learning is used to process tiles from whole slide images, then the need for pixel-level annotations is reduced, but the complexity of the inference system increases
Solution Approach 1:
The patent introduces an intermediary aggregation module that combines the predictions from multiple tile-level neural networks. This intermediary component handles the complexity of integrating multiple instances while keeping the individual tile processing relatively simple, thus reducing the need for complex pixel-level annotations while managing inference system complexity through modular design.
Data Source
AI summary
The present disclosure is directed to systems and methods for classifying biomedical images. A feature classifier may generate a plurality of tiles from a biomedical image. Each tile may correspond to a portion of the biomedical image. The feature classifier may select a subset of tiles from the plurality of tiles by applying an inference model. The subset of tiles may have highest scores. Each score may indicate a likelihood that the corresponding tile includes a feature indicative of the presence of the condition. The feature classifier may determine a classification result for the biomedical image by applying an aggregation model. The classification result may indicate whether the biomedical includes the presence or lack of the condition.