Hyperspectral-Deep-Learning Fusion for Rapid Differentiation of Adrenocortical Tumors and Pheochromocytomas

Accurate differentiation between adrenocortical tumors (ACTs) and pheochromocytomas (PHEOs) remains challenging in intraoperative pathology. This study proposes a hyperspectral-deep-learning fusion framework integrating spectral and spatial features for rapid, label-free diagnosis. Using Savitzky-Golay-filtered hyperspectral data (400-1000 nm) and a lightweight dual-branch convolutional network, the model achieved an AUC of 0.96 and an accuracy of 92.8%. Mutual information-based band selection and mid-level feature fusion enhanced efficiency, while Grad-CAM and spectral sensitivity analyses revealed interpretable wavelength-feature associations corresponding to lipid and hemoglobin absorption. The proposed system demonstrates strong potential for real-time, interpretable intraoperative discrimination between cortical and medullary adrenal tumors, supporting precision endocrine surgery.