---
library_name: peft
license: other
license_name: health-ai-developer-foundations
license_link: https://developers.google.com/health-ai-developer-foundations/terms
base_model: google/medgemma-4b-it
tags:
  - medgemma
  - lora
  - medical-ai
  - brain-mri
  - brain-tumor-classification
  - hai-def
  - neuroimaging
datasets:
  - sartajbhuvaji/Brain-Tumor-Classification
language:
  - en
pipeline_tag: image-text-to-text
---

# MedGemma Brain MRI LoRA

**Brain tumor classification adapter fine-tuned on Brain Tumor MRI dataset using MedGemma 4B.**

Classifies brain MRI images into 4 categories: glioma, meningioma, pituitary tumor, or no tumor.

## Model Details

| Property | Value |
|----------|-------|
| **Base Model** | [google/medgemma-4b-it](https://huggingface.co/google/medgemma-4b-it) |
| **Method** | LoRA (Low-Rank Adaptation) |
| **Task** | Multi-class brain tumor classification (4 classes) |
| **Modality** | Brain MRI |
| **Framework** | PyTorch + HuggingFace Transformers + PEFT |

## Training Dataset

**Brain Tumor MRI Classification** dataset — multi-source brain MRI collection for tumor type classification.

Fallback sources tried in order: `masoudnickparvar/brain-tumor-mri-dataset`, `AIOmarRehan/Brain_Tumor_MRI_Dataset`, `sartajbhuvaji/Brain-Tumor-Classification`

- **Train samples:** ~5,700 (85% split)
- **Validation samples:** ~1,000 (15% split)
- **Split strategy:** `train_test_split(test_size=0.15, seed=42)`

### Class Distribution

| Label | Description |
|-------|-------------|
| glioma | Malignant tumor from glial cells. Irregular, heterogeneous mass with surrounding edema. Most common primary malignant brain tumor. |
| meningioma | Typically benign tumor from the meninges. Well-defined, homogeneously enhancing extra-axial mass with dural tail sign. |
| pituitary | Adenoma from the pituitary gland in the sella turcica. May compress the optic chiasm causing visual field defects. |
| notumor | Normal brain MRI without intracranial mass, hemorrhage, or significant abnormality. |

## Training Configuration

### LoRA Parameters

| Parameter | Value |
|-----------|-------|
| Rank (r) | 16 |
| Alpha | 32 |
| Dropout | 0.05 |
| Target Modules | all-linear |
| Task Type | CAUSAL_LM |
| Trainable Params | 1.38B / 5.68B (24.3%) |

### Hyperparameters

| Parameter | Value |
|-----------|-------|
| Epochs | 1 |
| Per-device Batch Size | 1 |
| Gradient Accumulation Steps | 8 (effective batch = 8) |
| Learning Rate | 2e-4 |
| LR Scheduler | Linear with warmup |
| Warmup Ratio | 0.03 |
| Max Grad Norm | 0.3 |
| Precision | bfloat16 |
| Gradient Checkpointing | Enabled |
| Seed | 42 |

### Infrastructure

| Property | Value |
|----------|-------|
| GPU | NVIDIA L4 (24 GB VRAM) |
| Cloud Platform | [Modal](https://modal.com) serverless GPU |
| Training Time | ~30-45 minutes |
| Final Training Loss | 0.1026 |

## Prompt Format

**Input:**
> Analyze this brain MRI and classify the finding.

**Output:**
> This brain MRI shows **Meningioma**.
>
> Meningioma (typically benign tumor arising from the meninges. Appears as a well-defined, homogeneously enhancing extra-axial mass, often with a dural tail sign).

## Usage

```python
from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel
from PIL import Image

base_model_id = "google/medgemma-4b-it"
adapter_id = "efecelik/medgemma-brain-mri-lora"

processor = AutoProcessor.from_pretrained(base_model_id)
model = AutoModelForImageTextToText.from_pretrained(
    base_model_id, torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)

image = Image.open("brain_mri.jpg").convert("RGB")
messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Analyze this brain MRI and classify the finding."}
    ]}
]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt", images=[image]
).to(model.device)

output = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(output[0], skip_special_tokens=True))
```

## Intended Use

This adapter is part of the **MedVision AI** platform built for the [MedGemma Impact Challenge](https://www.kaggle.com/competitions/med-gemma-impact-challenge). It is designed for:

- **Medical education**: Helping students learn brain tumor identification on MRI
- **Clinical decision support**: Assisting radiologists with brain lesion characterization
- **Research**: Exploring fine-tuned medical VLMs for neuroimaging

## Limitations

- **Not for clinical diagnosis.** This model is for educational and research purposes only.
- **Limited tumor types:** Only classifies 4 categories. Many brain pathologies (abscess, stroke, MS) are not covered.
- **Single sequence:** Trained on individual MRI slices, not full 3D volumes or multi-sequence protocols.
- **Single epoch:** Trained for 1 epoch; further training may improve performance.

## Disclaimer

This model is for **educational and research purposes only**. It is NOT intended for clinical diagnosis or patient care decisions. Always consult qualified medical professionals for medical advice.