Dudda Nagaraju, Abrahamsen, Moradi, and Elschot at the Norwegian University of Science and Technology and St. Olavs Hospital submitted arXiv:2608.18731 on 19 August 2026. The paper is A Few Cases Are All You Need: An Empirical Study of Annotation-Efficient LoRA Fine-Tuning of MedSAM3. They freeze MedSAM3 and train LoRA adapters as organ- and modality-specific binary models for liver, kidney, spleen, gallbladder, and pancreas in CT and MRI. Few-shot sets of 1, 2, 5, and 10 cases come from the TotalSegmentator CT and MRI collections. Testing is the independent AMOS22 benchmark.
On AMOS22, 10-shot CT-trained LoRA takes gallbladder Dice from 0.0003 for TotalSegmentator and 0.0004 for MRSegmentator to 0.684 ± 0.028 (three runs). The same CT-trained adapters score 0.633 ± 0.034 on MRI gallbladder against 0.0004 for both specialists. Zero-shot MedSAM3 sits at 0.044 CT and 0.013 MRI on that organ. Liver CT is 0.928 ± 0.008 against MRSegmentator 0.959. Kidney CT is 0.898 ± 0.006 against 0.925. Pancreas CT is 0.711 ± 0.017, above TotalSegmentator 0.624 and below MRSegmentator 0.811. Training is about 3 to 5 GPU-hours per organ on one NVIDIA A40 48 GB, updating 18.5 M of 859 M parameters (rank 16). The preprint does not list a code URL for this study. Rebuild from the public MedSAM3 checkpoint and the paper protocol.
Frozen MedSAM3, LoRA on every encoder
MedSAM3 is a text-prompted SAM3 medical checkpoint (Liu et al., arXiv:2511.19046). Public code sits at github.com/Joey-S-Liu/MedSAM3 and LoRA weights at huggingface.co/lal-Joey/MedSAM3_v1. The NTNU study keeps those pretrained weights frozen. LoRA is applied to the vision, text, and geometry encoders, the DETR encoder and decoder, and the mask decoder. The organ name is the text prompt. The head predicts a binary mask. Loss is Dice plus focal. Optimiser is AdamW with weight decay, mixed precision, gradient clipping, linear warm-up, and cosine decay. The paper does not publish the learning rate.
Each adapter is one organ and one modality. That is ten abdominal models at the 10-shot operating point. The MedSAM2-Anatomy post is a frozen TotalSegmentator prior with three MedSAM2 tracks and a union. This paper is few-shot LoRA on MedSAM3.
Rank 16, 30 epochs, last checkpoint
Phase 1 used kidney at the 10-case budget to pick rank and epoch count. They compared r in {4, 8, 16} and watched a TotalSegmentator hold-out of 30 CT and 30 MRI cases every five epochs. That hold-out is never used for early stopping, learning-rate scheduling, or checkpoint selection. Rank 16 was selected. Kidney validation Dice at epoch 30 was 0.93 CT and 0.75 MRI. Liver and spleen curves under the same setting plateau by epoch 30. Gallbladder and pancreas were not in Phase 1.
Phase 2 locks r=16, 30 epochs, and the last checkpoint. No validation-based model selection. AMOS22 is disjoint from the TotalSegmentator train and hold-out splits. Ten-shot abdominal numbers are mean ± std over three independent runs. If you still mix the coefficient with the training loss, read the note on Dice versus Dice loss.
Gallbladder is the ugly cell on AMOS22
AMOS22 is 500 CT and 100 MRI scans with 15 abdominal organs (Ji et al., 2022). The paper reports Dice on that independent test set. Table 1, 10-shot LoRA versus the frozen specialists:
- Gallbladder CT: TotalSegmentator 0.0003, MRSegmentator 0.0004, LoRA 0.684 ± 0.028. MRI: both specialists 0.0004, CT-trained LoRA 0.633 ± 0.034, MRI-trained LoRA 0.593 ± 0.050.
- Liver CT: 0.951 / 0.959 / 0.928 ± 0.008. MRI: 0.921 / 0.956 / MRI-trained 0.927 ± 0.012.
- Kidney CT: 0.903 / 0.925 / 0.898 ± 0.006. MRI: 0.875 / 0.944 / 0.884 ± 0.019.
- Spleen CT: 0.924 / 0.952 / 0.898 ± 0.026. MRI: 0.712 / 0.935 / MRI-trained 0.845 ± 0.039. CT-trained LoRA on MRI spleen is 0.900 ± 0.021.
- Pancreas CT: 0.624 / 0.811 / 0.711 ± 0.017. MRI: 0.505 / 0.788 / MRI-trained 0.671 ± 0.017. CT-trained LoRA on MRI pancreas is 0.710 ± 0.014.
The authors put the specialist collapse on AMOS22 gallbladder down to domain shift, organ-size variability, and weak contrast. Zero-shot SAM3 and MedSAM3 never clear 0.05 Dice on gallbladder. The 0.68 number is a recovery against empty specialist masks. No working specialist gallbladder baseline is in the table. MRSegmentator still leads pancreas and MRI spleen. The paper reports Dice only. There is no HD95.
Mean Dice across the five organs rises 10.6% (CT) and 24.2% (MRI) from 1 to 5 cases, then 4.4% (CT) and 3.5% (MRI) from 5 to 10. nnU-Net at 10 cases is 0.11 Dice on MRI spleen and 0.39 on CT pancreas. nnU-Net full-pool sizes on the TotalSegmentator training side are 746 / 652 / 704 / 362 / 491 CT cases for liver, kidney, spleen, pancreas, and gallbladder, and 193 / 175 / 180 / 146 / 103 MRI cases. nnU-Net runs 500 epochs at 8.5 to 11 GPU-hours per organ on the same A40, with a fixed step count per epoch, so a 10-case budget does not get cheaper to train.
Cross-center WHS cardiac
Phase 3 moves off abdomen. CT adapters train on 10 cases from center A and test on 20 from center B. MRI adapters train on 10 from center E and test on 20 from centers C and D. Targets are left and right ventricle. TotalSegmentator CT reports LV 0.896 and RV 0.915. TotalSegmentator MRI and MRSegmentator have no cardiac labels in this comparison. 10-shot LoRA is 0.909 LV and 0.785 RV on CT, 0.820 LV and 0.637 RV on MRI. CARE-WHS 2025 winners trained on 86 cases sit at 0.943 / 0.928 CT and 0.935 / 0.914 MRI. RV MRI at 0.637 is the weak cell. The paper points at trabeculations, thin walls, and the cross-center split. Table 2 has no run-to-run std.
How this lands in a viewer
If you already hang AI in a DICOM viewer, treat this as a site-adapter recipe for an organ the public TotalSegmentator or MRSegmentator mask leaves empty, gallbladder on AMOS22 being the measured example. Keep the specialist CNN on liver, kidney, and spleen, and keep MRSegmentator on pancreas. Cardiac MRI is a label the two specialists do not offer here. 10-shot LoRA is 0.82 LV Dice; 0.64 RV sits well below the 86-case CARE-WHS RV number (0.914). RadYOLO is still the faster first-pass box. SAT3D is tumour plus uncertainty in Slicer. SLIP is click latency and undo. This paper is text-prompted MedSAM3 plus 10 labeled volumes and a rank-16 adapter.
Hardware in the paper is one A40. Adapters are binary and do not share a decoder. Rank and epoch were chosen on kidney, with liver and spleen as a check, then applied to gallbladder and pancreas. Ten cases is where their mean-Dice curve flattened. Retrospective AMOS22 overlap is not a device claim. The preprint is the source of record until a venue version exists. Rebuild from MedSAM3 and the protocol in arXiv:2608.18731. Do not wait on study weights. None were listed.
Sources
- Dudda Nagaraju, S., Abrahamsen, B. S., Moradi, A., Elschot, M. A Few Cases Are All You Need: An Empirical Study of Annotation-Efficient LoRA Fine-Tuning of MedSAM3. arXiv:2608.18731. Submitted 19 August 2026. https://arxiv.org/abs/2608.18731 (HTML: https://arxiv.org/html/2608.18731)
- Liu, A., Xue, R., Cao, X. R., Shen, Y., Lu, Y., Li, X., Chen, Q., Chen, J. MedSAM3: Delving into Segment Anything with Medical Concepts. arXiv:2511.19046. https://arxiv.org/abs/2511.19046. Code: https://github.com/Joey-S-Liu/MedSAM3. Weights: https://huggingface.co/lal-Joey/MedSAM3_v1
- Wasserthal, J., et al. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiol. Artif. Intell. 2023. DOI: 10.1148/ryai.230024
- Häntze, H., et al. Segmenting Whole-Body MRI and CT for Multiorgan Anatomic Structure Delineation. Radiol. Artif. Intell. 2025; 7(6):e240777. DOI: 10.1148/ryai.240777
- Ji, Y., et al. AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation. NeurIPS 2022. arXiv:2206.08023
- CARE-WHS 2025 whole-heart track: https://zmic.org.cn/care_2025/track3/