Medical imaging · 2025–present
LiPIVOT: liver and tumour segmentation across contrast-enhanced CT phases
Adapting a medical segmentation foundation model to outline the liver and liver tumours on CT across contrast phases and independent cohorts.
Qatar University · Ongoing research
Approach
Can fine-tuning a segmentation foundation model on liver CT outperform the model used as it is, and stay reliable on hospitals, scanners and contrast phases it was not trained on?
Liver tumours are assessed on CT scans taken at several times after a contrast injection, and each phase shows the liver and lesions differently. LiPIVOT starts from MedSAM2, a foundation model that segments 3D scans by propagating a mask from one prompted slice to its neighbours. The portal-venous phase is the anchor: the model is first adapted for the liver on that phase, then extended to tumours and to the other phases, and the phase-specific results are finally brought into a common space and combined.
- Detect the liver and prompt the model
- Segment through each CT volume
- Combine the contrast phases
LiverSAM: adapting MedSAM2 to portal-venous CT
Stage 1 · experiments in progress
The first stage fine-tunes MedSAM2 for the liver on about 600 portal-venous CT scans from eight public datasets. Each dataset is held out in turn, so every test measures performance on a cohort the model has not seen. Untouched MedSAM2 is compared with four fine-tuning settings that unfreeze progressively more of the network, from the mask decoder alone to the image encoder.
The untouched model is already a strong baseline, so the protocol is designed to make the comparison fair. Training follows the same slice-by-slice propagation used at test time and includes slices beyond the liver, so the model learns where the organ ends. Evaluation reports surface-distance measures and mask leakage outside the liver as well as overlap. A YOLO liver detector provides the starting box prompt for scans without reference masks.
Extending to tumours and other contrast phases
Stages 2 and 3 · data prepared
Next, a tumour model and phase-specific experts will be adapted using further public multiphase liver CT collections. Their segmentations will be mapped into the portal-venous space of each patient and fused into a consensus. An optional final stage will classify tumours from the combined phases.
My contribution
I lead the LiPIVOT pipeline: curating and quality-checking the eight-dataset CT collection, the LiverSAM training and leave-one-dataset-out evaluation protocol, automatic prompting, and the metrics used to compare fine-tuned and untouched models.
Study scope
This is ongoing work, so this page gives an overview only. Results will be added as the studies are published. Scores that use a box prompt drawn from the reference mask are reported separately from fully automatic results. The published liver MRI project is a separate study.