Medical Image Segmentation and Saliency for Explainable Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models in clinical settings lack explainability, with saliency maps being imprecise and difficult to interpret, hindering their adoption in medical image analysis.
Innovation Solution
Augmenting machine learning model outputs with a combination of image segmentation and saliency maps to provide anatomical context, enabling clearer interpretation of which features led to a prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If saliency maps are used to explain machine learning model predictions, then interpretability is improved, but precision and clarity of explanation deteriorate due to fuzzy and difficult-to-interpret results
Solution Approach 1:
The patent segments the medical image into multiple anatomical regions using a segmentation model, then generates separate saliency maps for each region. This segmentation approach transforms the single fuzzy saliency map into multiple precise regional explanations, allowing clinicians to clearly identify which specific anatomical structures contributed to the diagnosis while maintaining interpretability through region-specific visualizations.
2Reliability
If machine learning models are used for medical image analysis, then accuracy approaches or exceeds human experts, but explainability deteriorates creating a black-box effect
Solution Approach 1:
The patent introduces an intermediary explanation system that bridges the black-box ML model and the clinician. The system generates saliency maps that highlight important image regions and natural language explanations that describe the reasoning process. This intermediary layer maintains the high diagnostic accuracy of the ML model while providing transparent, interpretable explanations that clinicians can verify and trust.
Solution Approach 2:
The system provides feedback to clinicians by displaying both the ML model's diagnosis and the supporting evidence through saliency maps and textual explanations. This feedback loop allows clinicians to verify the model's reasoning, understand the basis for predictions, and maintain confidence in the system's high-accuracy diagnoses without being overwhelmed by the complexity of the underlying algorithms.
3Device complexity
If a single saliency map is generated for the entire image, then computational complexity is reduced, but information completeness deteriorates by indicating large swathes or regions from various different features
Solution Approach 1:
The patent divides the medical image into multiple anatomical segments and generates a separate saliency map for each segment. This segmentation strategy preserves detailed feature information by allowing each region to be analyzed independently, preventing the dilution of important diagnostic features that occurs in global saliency maps. The computational complexity increases moderately but remains manageable while significantly improving information completeness.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
According to an aspect, there is provided an apparatus for analysing a medical image, the apparatus comprising: a memory comprising instruction data representing a set of instructions; and a processor configured to communicate with the memory and to execute the set of instructions. The set of instructions, when executed by the processor, cause the processor to: i) use a first ML model to predict a label for the image; ii) obtain a saliency map indicating relevancy of different regions of the image to the first ML model when the first ML model predicted the label for the image; iii) obtain a segmentation of the image; and iv) determine an indication of a feature in the image that led to the label being predicted by the first ML model, using the segmentation and the saliency map in combination.