73: Explainable Disease Classification via Multi-Ultrasound Images using Graph Neural Networks

Thomas M. Menino Convention & Exhibition Center 

Graph Neural Networks (GNNs) have shown promise in computer-aided diagnosis. However, the lack of interpretability hinders their clinical adoption. We present an explanation framework that adapts Local Interpretable Model-agnostic Explanations (LIME) to graph-level prediction by perturbing each patient's image set. Instead of superpixels, we sample images subsets (subgraphs) and estimate image-level influence on the model. To generate informative perturbations, we use a two-stage Adaptive Class-Balanced Sampling scheme. Stage I utilizes random sampling, and Stage II employs class-biased sampling, drawing 85% of images from the minority class pool to ensure an informative design matrix. Ridge and elastic net classifiers trained with 10-fold cross-validation quantify conditional image influence, and Pearson correlation provides complementary marginal relationships. We incorporate bootstrap-based uncertainty quantification to report standard errors and confidence intervals. To demonstrate the efficacy of our approach, we apply our method on 135 fatty liver disease patients from MacKay Memorial Hospital in Taiwan, offering clinicians intuitive visual explanations.

explainable AI

graph neural networks

ultrasound imaging

interpretability

fatty liver

deep learning 

Section on Statistics in Imaging