erasmusmc-bigr/medgemma-4b-lirads-features-lora

MedGemma 4B — Liver Lesion LI-RADS LoRA Adapter

Fine-tuned LoRA adapter for MedGemma 4B IT (google/medgemma-4b-it) on binary classification of seven LI-RADS features in liver MRI for hepatocellular carcinoma (HCC) detection.

Developed at the Biomedical Imaging Group Rotterdam (BIGR), Erasmus MC, Rotterdam, Netherlands.

---

Model Details

Base model	google/medgemma-4b-it
Fine-tuning method	LoRA (PEFT)
LoRA rank / alpha	16 / 16
LoRA dropout	0.05
LoRA target modules	all linear layers
Task	Binary classification (present / absent) per LI-RADS feature
Input	Multi-sequence MRI patch + text prompt
Output	present_prob / absent_prob + binary label

---

Training Data

Split	Patients	Lesions	Feature Pairs
Train	53	106	523
Validation (checkpoint selection)	11	23	119
Internal test	12	23	114
Total (internal)	76	152	756

• Source: OpenSwiss — open-access multi-centre liver MRI dataset with HCC cases
• Sequences used: T1 native, T1 hepatic arterial, T1 portal venous, T1 delayed, T2, DWI (ADC)
• Features annotated: 7 LI-RADS major features (see below)
• Annotation: Radiologist-confirmed ground truth
• Ethics & consent: OpenSwiss is a publicly released, de-identified research dataset made available under an open data license; the original data collection was covered by the institutional ethical approvals and patient consent procedures of the contributing centres. No patient-identifiable information is contained in this repository or in the released model weights.

---

LI-RADS Features

1. Delayed phase capsule enhancement
2. Delayed phase washout appearance
3. Diffusion restriction
4. T1 native signal enhancement
5. T2 hyperintensity
6. Venous phase capsule enhancement
7. Venous phase washout appearance

---

Evaluation

This adapter was evaluated with five-fold patient-level cross-validation on the internal cohort (76 patients, 152 lesions from OpenSwissHCC and LiverHccSeg, followed by external validation on an independent institution. It was compared against prompting-only baselines , the matched fine-tuned model, and a soft-voting ensemble of the two fine-tuned 4B models.

External validation cohort: 107 cases were randomly selected (fixed seed) from the public LLD-MMRI dataset (Lou et al., 2025), sourced from Ningbo Medical Center Lihuili Hospital: hepatocellular carcinoma and hepatic cyst, liver metastases, intrahepatic cholangiocarcinoma, focal nodular hyperplasia and hepatic hemangioma, 107 cases were annotated independently by a single senior abdominal radiologist at Erasmus MC.

Results are reported in the associated manuscript, which is currently under peer review, and will be added here once it is published. For collaboration inquiries in the meantime, please contact the corresponding author.

Reference: Lou, M., Ying, H., Liu, X., Zhou, H.-Y., Zhang, Y., & Yu, Y. (2025). SDR-Former: A Siamese Dual-Resolution Transformer for Liver Lesion Classification Using 3D Multi-Phase Imaging. Neural Networks, 107228. Dataset: LLD-MMRI-Dataset

---

Usage

from peft import PeftModel
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

base = AutoModelForImageTextToText.from_pretrained(
    "google/medgemma-4b-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(
    base,
    "erasmusmc-bigr/medgemma-4b-liver-lesion-lora",
)
processor = AutoProcessor.from_pretrained(
    "erasmusmc-bigr/medgemma-4b-liver-lesion-lora"
)

Note: This model expects multi-sequence MRI input formatted with the prompt library from the original study. Standalone inference on arbitrary MRI images is not guaranteed without the preprocessing pipeline.

Note: google/medgemma-4b-it is a gated model on Hugging Face. You must request and be granted access to the base model from Google before you can load it.

Inference

Running this model requires the full inference pipeline from the original study:

• Multi-sequence MRI patch extraction (T1 native, arterial, portal venous, delayed, T2, DWI/ADC)
• Feature-specific prompt formatting (7 LI-RADS features)
• Output probability parsing (present_prob / absent_prob)

The complete pipeline is available at: https://gitlab.com/radiology/radiomics/lai/vision_language_models/liver-lesion-vlm

For enquiries, contact: Dr. Martijn Starmans, Assistant Professor, Erasmus MC m.starmans@erasmusmc.nl

---

Intended Use & Limitations

• Intended use: Research on automated LI-RADS feature detection in liver MRI.
• Not for clinical use: This model has not been validated for clinical decision-making.
• Dataset size: Trained on a small single-centre dataset (53 patients); performance on out-of-distribution data may vary.
• Annotation scope: Only the 7 LI-RADS features listed above were annotated; the model does not produce overall LI-RADS scores or diagnoses.

---

Data Availability

• Model weights (this repository): erasmusmc-bigr/medgemma-4b-liver-lesion-lora
• Training/internal test data: OpenSwiss, open-access multi-centre liver MRI dataset (Zenodo)
• External validation data: Images, lesion masks, and diagnosis labels from LLD-MMRI are publicly available under the dataset's own terms (official repository). LI-RADS feature-level annotations for the HCC subset used in this evaluation were generated internally at Erasmus MC and are available from the corresponding author upon reasonable request.
• Inference / preprocessing code: GitLab repository

No patient-identifiable data is included in any of the above resources.

---

Citation

Allahchim, A. (2026). Biomedical Imaging Group Rotterdam (BIGR), Erasmus MC. (Publication forthcoming — citation will be updated upon release.)

---

License

This adapter is derived from google/medgemma-4b-it, which is distributed under Google's Health AI Developer Foundations terms of use — not the standard Gemma Terms of Use. Use of this adapter, including access to the underlying base model weights, is subject to those terms.

google/medgemma-4b-it is a gated model on Hugging Face; users must request and be granted access from Google before they can load the base model to run this adapter.