Subimage Classification Probability Maps for Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine-learning techniques, particularly CNNs, struggle to effectively analyze large medical images and convey the rationale behind classifications due to their opaque nature and inability to visualize critical image regions, leading to clinical uncertainties.
Innovation Solution
The method involves decomposing large images into smaller subimages, applying CNNs for classification, and generating probability maps that combine overlapping subimage classifications to create a visualization that highlights all relevant regions and their associated probability levels, enhancing classification accuracy and clinical understanding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNNs are applied to large medical images directly, then classification accuracy may be improved, but computational complexity and memory resources required increase significantly
Solution Approach 1:
The patent divides large medical images into smaller subimage regions or tiles that can be processed individually by CNNs. This segmentation allows the system to handle large images that would otherwise exceed memory constraints while maintaining classification accuracy through aggregation of subimage probabilities.
2Loss of information
If CNNs process every pixel in the image, then complete image analysis is achieved, but the ability to convey rationale behind classification is lost due to black-box nature
Solution Approach 1:
By segmenting the image into subregions and processing them separately, the system can track which specific subimage regions contribute to the overall classification. This enables generation of probability maps that show which areas influenced the decision, providing interpretability while maintaining complete image analysis.
Solution Approach 2:
The patent introduces probability maps as an intermediary representation that bridges the black-box CNN classification and human interpretable rationale. These maps visualize the contribution of different image regions to the final classification, making the decision process transparent without sacrificing analytical completeness.
3Ease of operation
If Grad-CAM techniques are used to highlight important regions, then some interpretability is provided, but the visualization is too small and requires painstaking analysis of many images
Solution Approach 1:
The patent merges multiple Grad-CAM visualizations into a single comprehensive probability map that aggregates information across all subimage regions. This consolidation provides complete interpretability in one view rather than requiring analysis of many separate small images, dramatically reducing the time and effort needed to understand classification rationale.
4Productivity
If large high-resolution images are decomposed into small tiles for CNN processing, then computational feasibility is improved, but the ability to visualize classification basis is lost
Solution Approach 1:
The patent adds a new dimensional layer by creating probability maps that overlay spatial information back onto the full-image coordinate system. This allows the system to maintain computational efficiency through tile-based processing while recovering and visualizing spatial context through the probability map representation.
Data Source
AI summary
Large digital images are classified by analyzing them at a subimage level and assigning classification probabilities to the subimages; these may be combined into a classification probability for the entire image. Classification probabilities may be visualized across the image using probabilities computed for the subimages. This enables ready identification of the image subregions upon which classification is based as well as the classification scores or probabilities associated therewith. For example, a large source image—too large to be analyzed directly by a neural network—may be decomposed into smaller subimages such as square tiles, which are sifted based on a visual criterion. The visual criterion may be image entropy, density, background percentage, or other discriminator. A neural network produces tile-level classifications that are aggregated to classify the source image, and overlapping tiles are used to create a probability map showing subimage probabilities.