Automatic and semi-automatic segmentation and detection of hepatocellular carcinoma in CT: A multicenter retrospective study

Objectives Compare fully-automatic and semi-automatic HCC detection/segmentation in abdominal CT across centres, and develop a phase‑agnostic segmentation model. Methods We used three retrospective patient cohorts with HCC segmentations: one for resection eligibility, with 225 arterial (AP) and 228 portal venous (PV) CT (Beaujon), and two for transarterial chemoembolisation treatment eligibility, with 33/230 AP and 51/231 PV CT (HCCTACE/WAWTACE, respectively). We trained four individual phase-specific models on Beaujon and WAWTACE, and a phase-agnostic model on both datasets. We tested these models and MedSAM/MedSAM2 (two semi-automatic medical image segmentation foundation models) on held-out data from Beaujon and WAWTACE and on HCCTACE. Performance was assessed with Dice scores (DSC) and recall at intersection-over-union (IoU)>0.1. Statistical analysis was performed with multivariate analyses and post-hoc testing. Results All models had comparable DSC (p > 0.05), with semi-automatic models showing higher recall. Centre alignment training/testing improved DSC by 0.041 (p < 0.001). Per-case recall was high (lowest observed recall case =0.80) but per-lesion recall could be as low 0.57. For automatic models, the phase‑agnostic model matched phase‑specific models in performance (p>0.05) and its weights are made available. Semi-automatic model performance is sensitive to bounding box deformations. CT lesion-specific MedSAM2 performs comparably to MedSAM, while generalist MedSAM2 performs worse. Conclusion Automatic HCC detection in liver CT is promising, but identifying all lesions may require semi-automatic/manual approaches. We make a well-performing HCC segmentation model available for the wider community.