Discretized Feature Interpretability in Thermal Image AI Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI-enabled medical diagnosis tools lack interpretability, making it difficult for clinicians to understand the reasoning behind predictions, especially with complex, non-linear feature values from digital images, which affects trust and usability in clinical practices.
Innovation Solution
A system that processes numeric features from thermal images by using a machine learning prediction model to generate discrete values indicating the contribution of features towards a predicted class, enabling clearer interpretation through a mapping function and report generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If continuous and dynamic range feature values are extracted from medical images using AI-based analysis, then the accuracy and detail of the analysis improve, but the interpretability and understandability of the features deteriorate
Solution Approach 1:
The patent transforms continuous feature values into discrete grading scale values through parameter transformation. This changes the state of the feature data from continuous to discrete, making it more interpretable for clinicians while preserving the essential information needed for accurate diagnosis.
Solution Approach 2:
The patent introduces an intermediary processing layer that includes mapping functions and clustering algorithms. This intermediary transforms the complex continuous features into discrete, interpretable grades that serve as a bridge between the AI model's internal representations and the clinician's understanding.
2Reliability
If deep learning models are used to learn statistical patterns from training data, then the predictive performance improves, but the semantic meaning and interpretability of features deteriorate
Solution Approach 1:
The patent extracts discrete semantic information from the deep learning model's continuous feature space by applying mapping functions and clustering. This extraction process retrieves the essential semantic meaning of features while maintaining the predictive power learned from training data.
Solution Approach 2:
The patent applies parameter transformation to convert continuous feature values with learned statistical patterns into discrete grades with semantic meaning. This transformation preserves the predictive accuracy while restoring interpretability by assigning semantic labels to different feature value ranges.
3Adaptability or versatility
If continuous feature values with unbounded ranges are used for prediction, then the model's ability to capture nuanced variations improves, but the ease of explaining and deciding severity deteriorates
Solution Approach 1:
The patent transforms unbounded continuous feature ranges into bounded discrete grading scales. This parameter transformation maintains the model's ability to capture nuanced variations through multiple grade levels while providing a simplified framework for explaining and making clinical decisions.
Solution Approach 2:
The patent segments the continuous feature space into discrete intervals or grades. This segmentation divides the unbounded range into manageable segments, each representing a specific severity level, making it easier for clinicians to understand and act upon the results while preserving the nuanced information through multiple segments.
Data Source
AI summary
A system and method for processing values of a subset of numeric features to provide interpretability of the results of a machine learning model, by determining an extent of the contribution of the subset of features towards a predicted class by performing: (i) receiving the thermal image, (ii) obtaining a region of interest in the thermal image of the subject, (iii) extracting a plurality of numeric features associated with the region of interest of the thermal image, (iv) predicting a class, (v) estimate an extent of contribution of the subset of numeric features towards the decision of the first machine learning prediction model (M) and (vi) generate a report that includes a generated discrete values that determines the extent of contribution of the subset of numeric features towards the predicted class.