AI-Driven Quantitative Decision and Surgical Planning System for Liver Cancer

Background Precise tumor boundary definition on preoperative imaging is often unreliable, and regional disparities in healthcare resources limit personalized decision-making. Currently, no comprehensive AI system integrates imaging, pathology, staging, decision, and surgical planning for liver cancer.

Objectives 1: Develop a pathological-tumor-boundary evaluation system based on whole-slide pathology and CT registration. 2: Build models to predict tumor boundary and satellite micro-metastasis from CT images using deep learning and radiomics. 3: Integrate staging modules (Child-Pugh, ECOG-PS, CNLC/BCLC) and generate individualized treatment recommendations. 4: Create a surgical planning platform that calculates liver remnant volume, vascular invasion metrics, anatomical variants, and performs virtual resections. 5: Validate the system in a retrospective self-controlled multi-reader multi-case study across multiple centers.

Methods Tumor boundary \& micro-metastasis prediction: Register 3D whole-slide pathology of resected specimens to preoperative CT using multiplanar reconstruction; pathologists annotate tumor and satellite lesions; train deep-learning models to predict pathological boundaries and micro-metastasis regions. Validate in 100 cases with pathological CT comparison. Decision modules: Automatically compute Child-Pugh and ECOG-PS scores from labs and records; integrate with tumor metrics and PV invasion to achieve CNLC/BCLC staging and generate decision suggestions. Surgical planning: Calculate functional liver volume requirements per consensus, estimate standard liver volume (SLV), tumor-bearing segment volume, future liver remnant (FLR/SLV), flag unsafe resections; analyze vascular invasion level, length, and perfusion territory; detect portal/bile anatomical variants for injury warnings; perform virtual anatomical and non-anatomical resections with margin control and risk predictions. Clinical validation: Conduct a retrospective, fully-crossed multi-reader multi-case trial: randomized CT reading by surgeons/radiologists with and without software assistance, separated by washout periods. Evaluate primary endpoints: AFROC-AUC for lesion detection, LROC-AUC for HCC diagnosis. Secondary endpoints include sensitivity, specificity, ROC-AUC, diagnostic consistency (Kappa), size/count accuracy (\<5% error), staging concordance, and reading time.

Study population Adult participants with dynamic contrast-enhanced liver CT, including HCC-positive, HCC-negative lesion-positive, and lesion-negative cases. Exclude cases with incomplete liver imaging, heavy noise, prior liver resection, or unreadable CT.

Statistical analysis Compare AUCs between AI-assisted and manual reading; non-inferiority established if lower bound of 95% CI \> 0. Secondary metrics include sensitivity, specificity, consistency, reading time, and staging accuracy.