This study proposed an intelligent intraoperative diagnostic framework that combines hyperspectral imaging (HSI) with deep reinforcement learning to accurately differentiate hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC), the two main subtypes of primary liver cancer. To address the limitations of conventional imaging techniques and serum biomarkers, the authors constructed the first clinical HSI dataset of liver tumors (n = 131, spectral range 400-1000 nm). The proposed method integrates a 3D residual neural network (3D-ResNet) with a Proximal Policy Optimization (PPO)-based reinforcement learning algorithm, framing spectral band selection as a Markov decision process. An intraclass constrained cross-entropy loss further enhances class separability and compactness. Experimental results demonstrate a classification accuracy of 95%, outperforming traditional band selection approaches. This framework enables rapid, real-time tumor subtyping during surgery, addressing the critical clinical need for timely and accurate liver cancer diagnosis, and offers a promising tool for advancing precision oncology and improving intraoperative decision making.
Keywords: cholangiocarcinoma; hepatocellular carcinoma; hyperspectral imaging; reinforcement learning; three‐dimensional convolutional neural network.
© 2025 Wiley‐VCH GmbH.