Lightweight Spectral-MLP for Hyperspectral Differentiation of Intrahepatic Cholangiocarcinoma and Hepatocellular Carcinoma.

Accurate differentiation between intrahepatic cholangiocarcinoma (ICC) and hepatocellular carcinoma (HCC) remains challenging due to overlapping imaging features. We propose a hyperspectral imaging (HSI)-based diagnostic framework using a lightweight Spectral-MLP network for label-free tumor classification. Seventy-nine liver specimens were analyzed after spectral calibration, noise-band removal, and normalization within 450-900 nm. Spectral analysis revealed distinct differences in hemoglobin-related bands (540-580 nm) and lipid/water absorption regions (760-850 nm), corresponding to vascular and stromal variations between ICC and HCC. The proposed Spectral-MLP achieved 93.6% accuracy, 93.1% sensitivity, 94.0% specificity, and an AUC of 0.965, comparable to deeper models but with higher efficiency. These findings demonstrate that hyperspectral data combined with a simple neural network can effectively distinguish ICC from HCC and suggest that the lightweight Spectral-MLP framework has the potential to provide interpretable, real-time, and clinically deployable decision support for optical pathological diagnosis.