Property	Value
Base Model	google/medgemma-4b-it
Method	LoRA (Low-Rank Adaptation)
Task	Multi-label thoracic disease classification (15 labels)
Modality	Chest X-ray (frontal/PA view)
Framework	PyTorch + HuggingFace Transformers + PEFT

Training Dataset

NIH ChestX-ray14 — 112K frontal chest X-rays with 14 disease labels.

Reference: Wang et al. 2017, CVPR - "ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks"

Train samples: 10,000 (curated subset from 112K)
Validation samples: 1,000

Pathology Labels (15 classes)

Label	Description
No Finding	No acute cardiopulmonary abnormality
Atelectasis	Partial or complete lung collapse
Cardiomegaly	Enlarged heart (cardiothoracic ratio > 0.5)
Effusion	Fluid in the pleural space
Infiltration	Opacity suggesting infection or inflammation
Mass	Solid lesion > 3cm, requires malignancy evaluation
Nodule	Focal opacity < 3cm
Pneumonia	Infectious consolidation with air bronchograms
Pneumothorax	Air in pleural space causing lung collapse
Consolidation	Dense opacification replacing air with fluid/pus/cells
Edema	Fluid in lung interstitium/alveoli (often from heart failure)
Emphysema	Hyperinflation with flattened diaphragms
Fibrosis	Reticular opacities with volume loss
Pleural_Thickening	Increased pleural surface density
Hernia	Abdominal contents in thoracic cavity

Note: Multi-label classification — multiple pathologies can co-occur in a single image.

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 serverless GPU
Training Time	~60-90 minutes

Prompt Format

Input:

Analyze this chest X-ray and identify any findings.

Output:

This chest X-ray shows Pneumonia.

Pneumonia (infectious consolidation of lung parenchyma with air bronchograms).

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-chest-xray-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("chest_xray.jpg").convert("RGB")
messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Analyze this chest X-ray and identify any findings."}
    ]}
]

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. It is designed for:

Medical education: Helping students learn systematic chest X-ray interpretation
Clinical decision support: Assisting radiologists with thoracic disease screening
Research: Exploring fine-tuned medical VLMs for chest radiography

Limitations

Not for clinical diagnosis. This model is for educational and research purposes only.
Label noise: NIH ChestX-ray14 labels were NLP-extracted from reports and contain noise (~10-30% error rate depending on pathology).
Curated subset: Trained on 10K of 112K available images.
Single view: Trained on frontal views only. Lateral views not included.
Single epoch: Trained for 1 epoch; further training may improve performance.

Citation

bibtex

@inproceedings{wang2017chestx,
  title={Chestx-ray8: Hospital-scale chest x-ray database and benchmarks},
  author={Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and Bagheri, Mohammadhadi and Summers, Ronald M},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={2097--2106},
  year={2017}
}

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.