Tissue Classifier Using Multiple-Instance Learning for Biopsy Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In digital pathology, machine learning classifiers face challenges in accurately diagnosing cancerous tissues due to the abundance of healthy tissue, which can lead to reduced classification quality, as they randomly select image regions that resemble both cancerous and healthy tissues.
Innovation Solution
The training of a histological tissue classifier is separated into two parts: training a multiple-instance ROI classifier to minimize error over all regions of interest, and then training a tissue classifier based on actual ROI outputs, allowing for the aggregation of noisy decisions and eliminating the need for fine-grain labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning classifiers randomly select image regions for training, then the classification process is simple, but the classification accuracy deteriorates because healthy tissue regions resemble cancerous tissue
Solution Approach 1:
The patent segments the tissue image into multiple regions of interest (ROIs) and applies multiple-instance learning where each ROI is an instance within a bag representing the entire tissue. This segmentation allows the system to focus on discriminative regions while maintaining the overall tissue context, resolving the contradiction between accuracy and complexity by structuring the training data to highlight meaningful patterns without requiring complex preprocessing
Solution Approach 2:
The patent introduces an intermediary aggregation layer that combines predictions from multiple ROI classifiers to produce the final tissue-level classification. This intermediary mechanism allows the system to leverage simple ROI-level classifiers while achieving high accuracy through their collective decision, effectively mediating between simple local analysis and accurate global classification
2Ease of manufacture
If all tissue in a cancerous sample is assumed to be cancerous, then the labeling process is simplified, but the classification quality deteriorates due to the abundance of healthy tissue regions
Solution Approach 1:
The patent applies partial action by labeling only the entire tissue slide at the bag level rather than requiring fine-grained labeling of every cancerous region within the tissue. This partial labeling approach maintains ease of data preparation while the multiple-instance learning framework automatically identifies which ROIs contribute to the cancerous classification, resolving the contradiction between labeling simplicity and classification quality
3Measurement precision
If multiple-instance learning is applied with heuristic ROI selection, then the training becomes more accurate, but the device complexity increases due to additional processing requirements
Solution Approach 1:
The patent segments the large tissue image into smaller, manageable ROIs that can be processed independently by individual classifiers. This segmentation reduces the computational burden on each classifier while maintaining high accuracy through the collective decision of multiple classifiers, effectively managing the trade-off between precision and processing complexity
Solution Approach 2:
The system processes only a subset of ROIs that are most relevant to the classification decision, rather than requiring exhaustive analysis of every region. The aggregation mechanism focuses computational resources on discriminative regions while still achieving accurate tissue-level classification, balancing precision requirements with processing complexity
Data Source
AI summary
Disclosed is a computer implemented method for fully automated tissue diagnosis that trains a region of interest (ROI) classifier in a supervised manner, wherein labels are given only at a tissue level, the training using a multiple-instance learning variant of backpropagation, and trains a tissue classifier that uses the output of the ROI classifier. For a given tissue, the method finds ROIs, extracts feature vectors in each ROI, applies the ROI classifier to each feature vector thereby obtaining a set of probabilities, provides the probabilities to the tissue classifier and outputs a final diagnosis for the whole tissue.