CADx Lung Risk Score Decomposition for Clinical Explicability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Computer Aided Diagnosis (CADx) systems for lung cancer risk assessment, particularly those using deep neural networks, lack explicability, making it difficult for clinicians to understand and trust the predictive scores due to the complexity of the models and inability to assess the contribution of known predictors.
Innovation Solution
A method and system that decompose the risk score into contributions from known predictors, allowing clinicians to understand how the model arrived at the score by correlating features with specific predictors, using a neural network and a loss function that enforces non-zero contributions for known features to match clinical knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for lung cancer risk assessment, then predictive accuracy is improved, but explicability deteriorates
Solution Approach 1:
The patent segments the feature descriptor into multiple feature groups, where each group is associated with a specific known predictor (e.g., size, shape, texture). This segmentation allows the system to maintain the predictive power of deep neural networks while providing explicability by showing which known predictors contribute to the risk score. The feature groups act as intermediaries that bridge the gap between complex model internals and clinically interpretable features.
2Measurement precision
If complex machine learning models are used, then predictive performance is improved, but device complexity increases
Solution Approach 1:
The patent introduces feature groups as intermediary structures between the complex neural network and the clinical interpretation layer. Each feature group serves as a mediator that aggregates multiple features and associates them with known predictors. This intermediary structure allows the complex model to maintain its predictive performance while presenting a simplified, interpretable interface to clinicians through the mapping to known predictors.
3Loss of information
If features are decomposed into groups associated with known predictors, then explicability is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary grouping of features during the model training phase, organizing features into groups associated with known predictors before making predictions. This preliminary action allows the explicability structure to be established in advance, so that during inference, the system only needs to retrieve and aggregate pre-grouped features rather than performing complex decomposition calculations in real-time.
Data Source
AI summary
A method for providing a lung disease risk measure in a Computer Aided Diagnosis system is described. The method comprises the steps of: receiving an input comprising at least one input image showing all or part of the lungs of a patient; analysing the input to identify a feature descriptor comprised of at least one feature group, where each feature group comprises at least one feature computed from the input image; calculating a score explanation factor for each feature group, and an overall disease risk score for the patient's risk of lung disease from the feature descriptor; outputting the overall disease score risk score and a corresponding score explanation factor for each feature group. A computer aided diagnosis system is also described, along with a method for training a computer aided diagnosis system.