Abstract
The proliferative hepatocellular carcinoma (HCC) subtype is associated with aggressive disease and poor prognosis, yet it remains challenging to identify non-invasively. This retrospective multicenter study developed a multitask deep learning system for detecting proliferative HCC and predicting survival after intra-arterial therapy (IAT). Contrast-enhanced CT scans from two cohorts (surgical resection, n = 398; unresectable HCC receiving transarterial chemoembolization or hepatic arterial infusion chemotherapy, n = 1749) were analyzed using an nnUNet-based segmentation pipeline for liver and tumor delineation. A novel Prototype Mamba Net (PMN) architecture was created to extract imaging features indicative of proliferative biology. The model achieved AUCs of 0.825 (95% CI: 0.781-0.884) and 0.792 (95% CI: 0.732-0.841) on training and testing sets, respectively, for detecting proliferative HCC. Prognostic nomograms combining radiomic and clinical variables outperformed traditional staging systems (time-dependent AUC: 0.83-0.87; integrated Brier score: 0.12 versus 0.18-0.23, all P < 0.001). Among low-risk patients, no significant difference in survival was observed. However, for high-risk patients, HAIC showed a significant survival benefit compared to TACE (training and testing, P < 0.001). This non-invasive deep learning method enables preoperative identification of proliferative HCC and supports personalized IAT treatment choices, potentially improving outcomes in unresectable HCC.