medgemma-finetuned-cxr-reasoning

🧠 MedGEMMA Reasoning Model — Fine-tuned on CXR-10K

This is a fine-tuned version of google/medgemma-4b-it, trained on the CXR-10K Reasoning Dataset consisting of chest X-ray images paired with step-by-step clinical reasoning.

---

🩻 Task

Multimodal Clinical Reasoning:
Given a chest X-ray image, the model generates a step-by-step diagnostic reasoning path covering:

• Lung fields
• Cardiac size
• Mediastinal structures
• Surgical history
• Skeletal findings

---

🧪 Example Usage (Inference)

from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image
import torch

# Load model and processor
model = AutoModelForImageTextToText.from_pretrained("Manusinhh/medgemma-finetuned-cxr-reasoning")
processor = AutoProcessor.from_pretrained("google/medgemma-4b-it")

# Load image
image = Image.open("example.png").convert("RGB")

# Create prompt
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Analyze this medical image and provide step-by-step findings."}
        ]
    }
]

# Tokenize and generate
inputs = processor.apply_chat_template(messages, tokenize=True, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=300)

print(processor.decode(output[0], skip_special_tokens=True))

---

📊 Training Details

• Base model: google/medgemma-4b-it
• LoRA Fine-tuning: Used peft with low-rank adapters
• Training set: 10k chest X-ray samples with reasoning steps
• Frameworks: HuggingFace Transformers, TRL, PEFT, DeepSpeed

---

📚 Dataset Attribution

Training data derived from: CXR-10K Reasoning Dataset Built upon: itsanmolgupta/mimic-cxr-dataset-10k

Base dataset: MIMIC-CXR by MIT LCP

Johnson AE, Pollard TJ, Berkowitz SJ, et al. Scientific Data. 2019;6:317. https://doi.org/10.1038/s41597-019-0322-0