Counterfactual Medical Image Explanations Using Morphological Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing explainable AI techniques in medical imaging fail to clearly elucidate how structural and functional patterns in medical images influence AI predictions, limiting the interpretability and trustworthiness of AI decision-making processes.
Innovation Solution
A method involving an AI predictive model that generates counterfactual explanations by identifying a source item with shared structural features but different prediction labels, masking a target item's morphological segment, and adding the source item's segment to create a recombined item, which is then analyzed to determine a new prediction label, indicating the morphological segment's influence on the prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional explainable AI techniques (feature visualization, saliency maps, gradient-based techniques) are used to provide insights into model decision-making, then the model's internal workings are partially revealed, but the explanation remains insufficient for clearly understanding how structural and functional patterns in medical images influence AI predictions
Solution Approach 1:
The patent segments the medical image into multiple morphological segments and systematically replaces each segment with corresponding segments from source images having different prediction labels. This segmentation approach enables precise identification of which specific morphological features drive the AI prediction, providing complete explanatory information about structure-prediction relationships without unnecessary complexity
Solution Approach 2:
The patent creates counterfactual images by copying morphological segments from source images and pasting them into target images. This copying mechanism allows direct visual comparison between original and modified images, making the explanation intuitive and easy to understand while maintaining complete information about how structural patterns influence predictions
2Reliability
If counterfactual images are generated by replacing morphological segments to change prediction labels, then interpretability is enhanced, but the process requires identifying source items with shared structural features but different predictions
Solution Approach 1:
The patent introduces source images as intermediary elements that share structural features with target images but have different prediction labels. These source images serve as mediators to isolate and identify the specific morphological features that drive prediction differences, enhancing interpretability while managing process complexity through systematic comparison
Solution Approach 2:
The patent applies local quality by focusing on specific morphological segments rather than analyzing entire images. Each morphological segment is individually replaced and evaluated to determine its specific contribution to the prediction, providing localized explanatory power that enhances overall interpretability without requiring complex global analysis
Data Source
AI summary
Provided are techniques for training and using machine learning models to provide counterfactual explanations of predictions. An Artificial Intelligence (AI) predictive model is trained. The AI predictive model is used to generate a prediction label for each item of a plurality of input items. A target item with an initial prediction label. For a morphological segment, a source item is identified from the plurality of input items, where the source item shares common structural features with the target item and has a different prediction label. A recombined item is generated by: masking the morphological segment in the target item and adding the morphological segment of the source item. The AI predictive model is used to generate a new prediction label for the recombined item. It is determined that the new prediction label is different from the initial prediction label and that the recombined item is a counterfactual item.