erjui/medgemma-4b-srrg-impression

medgemma-4b: Structured Radiology Report Generation (Impression)

This model is a fine-tuned version of google/medgemma-4b-it for generating the IMPRESSION section of structured chest X-ray radiology reports. It was trained using LoRA (Low-Rank Adaptation) on the csrrg_ift_dataset containing instruction-following examples from MIMIC-CXR and CheXpert+ datasets.

Model Description

This model performs Structured Radiology Report Generation (SRRG) for chest X-rays, specifically generating concise impression sections that summarize key clinical findings, differential diagnoses, and recommendations.

Key characteristics:

• Generates the IMPRESSION section of radiology reports
• Trained on single chest X-ray examinations
• Produces clinically relevant summaries and conclusions
• Fine-tuned with LoRA for parameter-efficient adaptation

Intended Use

Primary Use Cases

• Research on automated radiology report generation
• Development of clinical decision support systems
• Medical AI and multimodal model research
• Educational tools for radiology training

Intended Users

• Medical AI researchers
• Healthcare technology developers
• Clinical informatics specialists
• Radiology departments (research use only)

Out-of-Scope Use

• NOT intended for clinical diagnosis without physician review
• Should not replace human radiologists in clinical practice
• Requires validation before any clinical deployment

Training Details

Training Data

• Dataset: csrrg_ift_dataset (srrg_ift_dataset_impression subset)
• Training samples: ~405,971 instruction-following examples
• Data sources: MIMIC-CXR and CheXpert+ chest X-ray datasets
• Task format: Instruction fine-tuning with system-user-assistant conversations

Training Procedure

Fine-tuning method: LoRA (Low-Rank Adaptation)

LoRA Configuration:

• Rank (r): 32
• Alpha: 64
• Dropout: 0.1
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training hyperparameters:

• Learning rate: 2e-4
• Batch size: 4 per device
• Gradient accumulation steps: 32 (effective batch size: 128)
• Epochs: 1
• Optimizer: AdamW
• Learning rate scheduler: Cosine with 3% warmup
• Precision: bfloat16
• Attention implementation: Flash Attention 2
• Max sequence length: 2048
• Max images per sample: 1

Hardware:

• GPU: NVIDIA H100
• Training framework: HuggingFace Transformers + PEFT

Usage

Loading the Model

from transformers import AutoProcessor, AutoModelForVision2Seq
from PIL import Image
import torch

# Load model and processor
model_name = "erjui/medgemma-4b-srrg-impression"
model = AutoModelForVision2Seq.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("google/medgemma-4b-it", trust_remote_code=True)

# Load chest X-ray image (single image for SRRG)
image = Image.open("chest_xray.jpg")

# Prepare input
messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are an expert radiologist."}]
    },
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Analyze the chest X-ray images and write the IMPRESSION section of a radiology report. Provide a concise clinical summary and diagnosis based on the imaging findings."},
            {"type": "image"}
        ]
    }
]

# Process and generate (max_images_per_sample: 1)
inputs = processor(images=image, text=messages, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
generated_text = processor.decode(outputs[0], skip_special_tokens=True)

print(generated_text)

Expected Output Format

IMPRESSION:
1. Right apical rounded opacity concerning for infection or malignancy.
2. Recommend repeat dedicated AP and lateral chest radiograph, or CT for further evaluation.

Citation

If you use this model, please cite:

@article{kang2025automated,
  title={Automated Structured Radiology Report Generation with Rich Clinical Context},
  author={Kang, Seongjae and Lee, Dong Bok and Jung, Juho and Kim, Dongseop and Kim, Won Hwa and Joo, Sunghoon},
  journal={arXiv preprint arXiv:2510.00428},
  year={2025}
}

Also cite the base model:

@article{sellergren2025medgemma,
  title={Medgemma technical report},
  author={Sellergren, Andrew and Kazemzadeh, Sahar and Jaroensri, Tiam and Kiraly, Atilla and Traverse, Madeleine and Kohlberger, Timo and Xu, Shawn and Jamil, Fayaz and Hughes, C{\'\i}an and Lau, Charles and others},
  journal={arXiv preprint arXiv:2507.05201},
  year={2025}
}

Model Card Authors

Seongjae Kang (erjui)

Model Card Contact

For questions or issues, please open an issue on the model repository.