Deep Learning on Hyperspectral Pathology for Recurrence Prediction in Clear Cell Renal Cell Carcinoma
Xulei Wang, Wenshi Tian, Yihan Zhao, Yihui He, Zhengyang Zhang, Xiaobo Shao, Yunchao Wang, Jianning Wang · Journal of Biophotonics · 2025
BACKGROUND: Clear cell renal cell carcinoma (ccRCC), the most common aggressive renal cancer subtype, shows marked heterogeneity that hinders recurrence prediction. OBJECTIVE: To evaluate hyperspectral pathology imaging (HSI) with deep learning for individualized recurrence risk prediction in ccRCC. METHODS: Slides from 48 patients with ccRCC were imaged using a 400-1000 nm hyperspectral microscope. Spectral data were preprocessed, and a dual-branch network (HSI-FusionNet) extracted spatial and spectral features via 2D and 1D convolution, followed by gated fusion and multiple instance learning (MIL) for patient-level prediction. RESULTS: HSI-FusionNet achieved strong test performance (area under the receiver operating characteristic curve [AUC] = 0.912; sensitivity = 0.881; specificity = 0.846), outperforming ResNet-50, 1D-Convolutional Neural Network (CNN), and a 1D-Transformer. Recurrence-related spectral bands (530-580 and 830-900 nm) reflected hemoglobin and lipid-collagen differences. CONCLUSION: HSI with deep learning accurately identifies recurrent ccRCC and reveals molecular-metabolic signatures, supporting precision postoperative risk stratification.