🧠 brats_medgemma_light

brats_medgemma_light is the merged model derived from the fine-tuning experiments in
Jesteban247/medgemma-brats-experiments.

It is based on unsloth/medgemma-4b-it and represents the r1_alpha2_epochs1 configuration —
the lightweight variant fine-tuned on the BraTS and TextBraTS datasets for brain MRI interpretation and radiology text generation.

⚙️ Configuration Details

Parameter	Value
LoRA rank (r)	1
LoRA alpha (α)	2
Epochs	1
Trainable parameters	~2.4M (≈0.06%)
Total model size	4B parameters
Fine-tuning type	Full merge after LoRA training

📊 Quantitative Summary

Metric	ROUGE-1	ROUGE-2	ROUGE-L	Avg ROUGE
r1_alpha2_epochs1	0.5539	0.2765	0.3985	0.4096

This configuration serves as a minimal fine-tuning baseline in the experiment set,
balancing broad generalization with initial domain adaptation to medical language.

🧠 Qualitative Example

🖼️ Image:

Ground truth:

“The lesion area is in the right frontal and parietal lobes with a mixed pattern of high and low signals with speckled high signal regions. Edema is mainly observed in the right parietal lobe, partially extending to the frontal lobe, presenting as high signal, indicating significant tissue swelling around the lesion. Necrosis is within the lesions of the right parietal and frontal lobes, appearing as mixed, with alternating high and low signal regions. Ventricular compression is seen in the lateral ventricles with significant compressive effects on the brain tissue and ventricles.”

Model Output (r1_alpha2_epochs1):

“The MRI slice demonstrates: The lesion area is in the right cerebral hemisphere, mainly concentrated in the right frontal and parietal lobes with mixed signals... Edema is observed... Ventricular compression is observed with the right lateral ventricle being compressed.”

This output captures core clinical elements with concise phrasing but shows limited depth in signal pattern description and necrosis details compared to heavier LoRA configurations.

💬 Notes

This merged version reflects minimal domain adaptation due to fine-tuning on BraTS and TextBraTS with low-rank LoRA,
resulting in better generalization to non-medical tasks at the cost of shallower specialization in radiology-specific terminology.

Such trade-offs are typical in lightweight fine-tuning on small, domain-specific datasets,
and align with observations discussed in the parent project
➡️ Jesteban247/medgemma-brats-experiments