PROVEN: Hierarchical feature consistency and dual-prototype learning for interpretable medical diagnosis

The interpretability of models is crucial in high-stakes fields such as healthcare, as it allows researchers and physicians to understand the rationale behind decisions, thereby enhancing diagnostic credibility and reliability. Although prototype learning methods have shown great potential in providing interpretability, designing a network that achieves both high classification accuracy and medical interpretability in diagnostic tasks remains a significant challenge. Particularly when dealing with medical images where inter-class differences are subtle and complex, it is essential to extract rich, multi-scale features and obtain prototypes that intuitively reflect key lesion characteristics. To address these challenges, this paper proposes PRototype-Oriented Visual Explanation Network (PROVEN), a self-explanatory medical image classification method. This method effectively alleviates the issues of prototype redundancy and insufficient representativeness in traditional prototype learning by integrating two complementary types of prototypes: “Support Prototypes”, which focus on learning class boundaries to handle ambiguous samples, and “Trivial Prototypes”, which capture core class features to facilitate efficient classification. Furthermore, the model introduces a novel loss function—the Shallow-Deep Feature Consistency Loss (SDFCL)—to enhance the interaction between shallow and deep features, ensuring that local details are effectively propagated into deep representations. Extensive experiments were conducted on four public medical image datasets, including two cervical cytology datasets (SIPaKMeD and Hussian LBC) and two skin lesion datasets (PH2D7 and HAM10000), containing more than 16,000 images in total. Following the standard evaluation protocols of these benchmarks, PROVEN consistently outperforms state-of-the-art CNN-, Transformer-, and prototype-based methods. Compared with the best-performing prototype-based approaches, PROVEN achieves accuracy improvements of 5.80% on PH2D7 and 7.73% on HAM10000. These results demonstrate that PROVEN provides stronger inherent interpretability while achieving superior classification performance.