efecelik/medgemma-musculoskeletal-lora

MedGemma Musculoskeletal LoRA

Bone fracture detection adapter fine-tuned on fracture X-ray datasets using MedGemma 4B.

Binary classification of musculoskeletal X-rays: fractured vs. non-fractured, with optional body-part-specific context.

Model Details

Property	Value
Base Model	google/medgemma-4b-it
Method	LoRA (Low-Rank Adaptation)
Task	Binary fracture classification
Modality	Musculoskeletal X-ray
Framework	PyTorch + HuggingFace Transformers + PEFT

Training Dataset

Bone Fracture Detection — ~8.8K musculoskeletal X-ray images for binary fracture classification.

• Train samples: 8,000 (capped)
• Validation samples: 800
• Split strategy: train_test_split(test_size=0.15, seed=42)

Class Labels

Label	Description
fractured	Visible fracture line indicating cortical bone disruption. Pattern, location, and alignment should be characterized.
non-fractured	Intact cortical bone margins. No fracture line, displacement, or callus formation. Normal bone density and alignment.

Body Regions Covered

Hand/wrist, shoulder, elbow, knee, ankle, hip, forearm, humerus, finger, leg/tibia-fibula

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	~30-45 minutes

Prompt Format

When body part information is available:

Input:

Analyze this hand/wrist X-ray for fracture.

Output:

Fracture detected.

There is a visible fracture line indicating disruption of cortical bone continuity. The fracture pattern, location, and alignment should be further characterized for treatment planning.

When body part is unknown:

Input:

Analyze this musculoskeletal X-ray for fracture.

Usage

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

base_model_id = "google/medgemma-4b-it"
adapter_id = "efecelik/medgemma-musculoskeletal-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("wrist_xray.jpg").convert("RGB")
messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Analyze this wrist X-ray for fracture."}
    ]}
]

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 fracture identification on X-rays
• Clinical decision support: Assisting emergency physicians with fracture screening
• Research: Exploring fine-tuned medical VLMs for musculoskeletal imaging

Limitations

• Not for clinical diagnosis. This model is for educational and research purposes only.
• Binary only: Classifies fracture vs. no-fracture. Does not characterize fracture type (comminuted, spiral, greenstick, etc.).
• Subtle fractures: May miss occult fractures (stress fractures, non-displaced fractures) that are challenging even for radiologists.
• Single epoch: Trained for 1 epoch; further training may improve performance.
• Mixed body regions: Performance may vary across different anatomical regions.

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.