Conformal Prediction for Explainable AI and Lesion Detection in 3D Cranial Ultrasound

Punctate white matter lesions (PWML) are the most common white matter injuries observed in preterm neonates. With deep learning techniques, automatic detection of these lesions through cranial ultrasound imaging could provide doctors with valuable diagnostic support, at a lower cost than MRI. However, a prerequisite for trust in deep learning is uncertainty quantification, in order to obtain interpretable models that can be efficiently deployed in clinical settings. Conformal prediction (CP) is both a method of uncertainty quantification and a method of classifying instances, where uncertainty is conveyed by predicting sets of potential classes rather than single predictions. Using this CP framework in conjunction with an anomaly detection model, we can be deterministically confident in labeling a new test input as containing PWML or not. Experimental results demonstrate the effectiveness of our method in detecting PWML in ultrasound images, achieving a 5% improvement in precision compared to the best published models, while also providing uncertainty quantification for better interpretability.