Supervisor: Prof. Gemma Piella Fenoy
Institute: Universitat Pompeu Fabra (UPF), Barcelona, Spain
Duration: Oct 2025 – Present
This repository contains the complete workflow for developing a mobile-deployable liver boundary segmentation system for point-of-care imaging using RGB smartphone cameras.
The pipeline includes:
- SAM fine-tuning for precise liver segmentation
- Lightweight liver localization using TFLite
- Domain adaptation using color-calibrated data
- Ongoing YOLO-based detection with anatomical priors for robust localization
Early liver disease diagnosis often relies on resource-intensive imaging (CT, MRI).
This work enables low-cost, accessible liver boundary detection using ** consumer-grade cameras**, improving:
- Remote screening
- Telehealth workflows
- Biomedical imaging accessibility in low-resource regions
| Component | Method | Dataset | Performance | Status |
|---|---|---|---|---|
| Segmentation | Fine-tuned SAM-Vit-Base | Unseen validation | 0.90 Dice | Completed |
| Segmentation — Domain Adapted | SAM + Color-calibrated dataset | Adapted validation | 0.91 Dice | Completed |
| Liver Localization | Lightweight TFLite CNN | Smartphone RGB | 0.70 Dice (bbox) | Completed |
| Detection + Spatial Priors | YOLO-based anatomical guidance | WIP | — | In Progress |
⚡ Domain adaptation significantly improved edge precision on dark-skin and low-illumination samples.
notebooks/
finetuning.ipynb SAM fine-tuning for liver boundary segmentation
inference.ipynb Running segmentation/localization inference
Open the notebooks in Jupyter or Google Colab:
notebooks/finetuning.ipynb— fine-tune SAM (ViT-Base) on the liver segmentation dataset.notebooks/inference.ipynb— run inference with the fine-tuned model on new images.
A GPU runtime is recommended for both fine-tuning and inference.