A reinforcement learning-enhanced interpretable prototype deep network for predicting pulmonary nodule malignancy

Prototype-based deep networks classify images by comparing them with learned, semantically meaningful visual patterns (prototypes), enabling case-based reasoning and interpretable decision-making, that are critical for high-stakes medical diagnosis. However, these models face a trade-off between interpretability and predictive performance. First, medical imaging data (e.g., CT scans) consist of sequential slices, in which lesion radiological signs may vary across slices due to surrounding tissue interference. Existing prototype networks typically operate on individual slices, degrading their ability to consistently capture stable radiological patterns. Second, subtle and radiologically overlapping morphological differences between lesion types challenge current methods that rely on distances to a limited set of explicit prototypes, leading to insufficient preservation of discriminative information and reduced capacity to model complex visual patterns. To address these challenges, we propose RIPPLE-Net, a reinforcement learning-enhanced interpretable prototype network for pulmonary nodule malignancy prediction. It incorporates a reinforcement learning agent that mimics the sequential reading behavior of radiologists, dynamically refining prototype-based predictions by leveraging contextual information from adjacent slices. Furthermore, latent discriminative embeddings orthogonal to the predefined prototypes are learned in an unsupervised manner to capture subtle and radiologically overlapping morphological variations across lesion types. External validation on a real-world thoracic CT dataset shows that RIPPLE-Net improves AUC by 1.7% over three state-of-the-art prototype networks and by 2.3% over five black-box deep learning models, including both 2D and 3D sequence-based architectures. As the first framework to integrate reinforcement learning into prototype networks, RIPPLE-Net enhances diagnostic performance while preserving interpretability in diagnosis based on medical images.