Hepatocellular carcinoma (HCC) is the most common primary liver cancer and the third leading cause of cancer-related mortality worldwide, with early detection improving survival from <20% up to >70%.[1,2] Abdominal computed tomography (CT) is one of the primary imaging modalities for HCC diagnosis, with the standardized Liver Imaging Reporting and Data System (LI-RADS) defining specific diagnostic criteria that directly impact clinical decision-making and downstream treatment pathways.[3,4] Specifically, LI-RADS category 5 (LR-5) is considered “definitely HCC” on a scale from LR-1 through LR-5, and does not require further tissue biopsy for diagnosis [5-7] – a fully radiology imaging-based diagnostic pathway, where artificial intelligence (AI) tools can promote early diagnosis and affect patient outcomes.[8] Compared to other simpler medical imaging classification tasks, HCC diagnosis on abdominal CT presents a unique challenge based on established clinical LI-RADS criteria: The need for multi-phase acquisition requirements with different enhancement patterns (i.e. for arterial and washout assessment). Specifically, LI-RADS characterization depends on the presence of four imaging features: 1) Arterial phase hyperenhancement (APHE), where the lesion appears brighter than the surrounding liver tissue in the early arterial phase and serves as the primary entry point of HCC evaluation; 2) Size of the lesion; 3) Non-peripheral washout, where the lesion becomes darker than the liver in the portal venous or delayed phases; and 4) Enhancing capsule, where a thin, peripheral rim remains enhanced in the portal venous or delayed phases due to contrast retention. From a technical standpoint, these clinical requirements translate into distinct machine learning challenges: effective LI-RADS characterization for HCC requires spatio-temporal reasoning across contrast phases to capture dynamic enhancement patterns, robust multi-task learning architectures capable of modeling phase-specific lesion characteristics and cross-phase feature interaction, handling substantial class imbalance between LI-RADS categories, and strong out-of-distribution generalization across heterogeneous acquisition protocols. Dataset: To potentially train AI tools to automate HCC characterization on CT, public multi-phase liver CT datasets exist, however have historically been fragmented, heterogeneous, and limited in size. We have selectively harmonized four public CT datasets [9-13] into a unified, analysis-ready resource for HCC AI research. The dataset consists of 668 cases, including: 83 normal livers with no liver lesions, and 585 livers with HCC lesions. All HCC lesions were re-annotated with clinical LI-RADS categories and voxel-level segmentation masks were created for non-rim APHE, non-peripheral washout, and enhancing capsule by three board-certified abdominal radiologists (F.D., B.L., J.W.). Task: We propose a challenge focused on advancing AI-based characterization of HCC lesions in multi-phase abdominal CT scans using clinical imaging-based LI-RADS categories. Additionally, all submissions will be evaluated on a held-out private institutional test set to assess generalizability. This challenge aims to provide a standardized benchmark for clinically grounded liver lesion AI and drive methodological advances in multi-phase feature integration, multi-task learning for hierarchical diagnostic criteria, and domain generalization under varied real-world protocol heterogeneity – ultimately accelerating translation of AI tools into clinical workflows, where early, accurate HCC characterization can meaningfully improve patient outcomes.