New interpretable AI framework improves tumor classification with explainable imaging signatures

A team led by Chengkun Sun and collaborators at the University of Florida Health and other institutions has introduced a unified framework for interpretable imaging signature discovery in tumor classification. The method combines deep learning-based tumor segmentation, Grad-CAM attention mapping, and traditional radiomic analysis. First, a robust segmentation model delineates tumors. Then, Grad-CAM identifies diagnostically important regions, and a mutual information-based adaptive thresholding strategy extracts patient-specific signatures. These signatures are validated by feeding them into a downstream deep learning classifier, while radiomic features from the regions are interpreted using SHAP to pinpoint the most discriminative biomarkers. This pipeline directly addresses the 'black box' problem that has hindered clinical adoption of deep learning in radiology.

The framework was evaluated on four datasets: the public BUSI breast ultrasound, KiTS renal CT, BraTS brain tumor datasets, and a private UF Health renal CT cohort. Compared to conventional whole-tumor radiomics, the signature-based approach achieved improved discriminative performance while offering far greater biological interpretability. The key innovation is converting deep learning attention into reproducible quantitative imaging biomarkers that clinicians can understand and trust. By providing transparent, patient-specific signatures rather than opaque model decisions, this work represents a significant step toward non-invasive tumor characterization and imaging biomarker discovery. The authors have made the code and datasets available on arXiv, inviting further validation and clinical translation.