MERA: Multimodal and Multiscale Self-Explanatory Model with Considerably Reduced Annotation for Lung Nodule Diagnosis

Reduced annotation burden improves practical deployment: MERA achieves diagnostic accuracy comparable to fully-annotated systems using only 1% labeled data, addressing a major barrier to implementing AI in healthcare where expert annotation is expensive and time-consuming.

Multiple explanation levels enhance clinical trustworthiness: The system provides explanations at four different levels (global clustering patterns, similar case comparisons, visual attention maps, and specific nodule attributes), allowing clinicians to understand why the model made a diagnosis rather than treating it as a "black box."

Self-supervised learning enables learning from unlabeled data: By using unsupervised feature extraction through Vision Transformers before applying labeled data, MERA demonstrates that AI systems can learn meaningful patterns from the vast amounts of unlabeled medical images already available in hospitals.

Explainability is built into the model, not added afterward: Rather than using post-hoc explanation techniques applied after diagnosis, MERA's architecture inherently generates explanations, suggesting that responsible AI design should integrate interpretability from the ground up rather than as an afterthought.

Comprehensive evaluation on public datasets supports reproducibility: Testing on the public LIDC dataset (rather than proprietary data) enables independent verification of claims and facilitates broader adoption by other researchers and institutions.