medgemma-dsa-lora

Model Card: MedGemma DSA LoRA Adapter (ds-lora)

Model Details

Model Description

This repository contains a LoRA adapter fine-tuned for Digital Subtraction Angiography (DSA) vessel patency triage.

The adapter is intended to improve sensitivity for blocked/occluded vessel patterns compared with naive base-model prompting on heavily imbalanced data.

• Developed by: DataRabbit hackathon team
• Base model: google/medgemma-1.5-4b-it
• Model type: Vision-language LoRA adapter (PEFT)
• Primary task: Binary DSA support classification (patent vs blocked)
• Language: English (prompts/output)
• License: Follows base model terms and license restrictions
• Finetuned from model: google/medgemma-1.5-4b-it

Model Sources

• Project repository: local project (proj-medgemma-challenge-2026)
• Base model: https://huggingface.co/google/medgemma-1.5-4b-it
• Competition context: https://www.kaggle.com/competitions/med-gemma-impact-challenge/overview

Intended Uses

Direct Use

• Rapid DSA image triage support in a controlled evaluation pipeline.
• Experimental decision-support where output is always reviewed by a clinician.

Downstream Use

• Can be routed only for DSA/XA series in a multi-stage pipeline (e.g., PACS worker).
• Can be merged into a standalone model for simpler deployment.

Out-of-Scope Use

• Not for autonomous diagnosis.
• Not for emergency triage without human oversight.
• Not validated for non-DSA modalities (CT, MR, US, X-ray).
• Not intended for pediatric/special populations without dedicated validation.

Bias, Risks, and Limitations

• Performance can degrade under domain shift (scanner/protocol/institution changes).
• Small and previously imbalanced blocked-vessel class is a key risk.
• Model outputs can be verbose or overconfident; structured post-processing is recommended.
• False negatives remain clinically high-risk.

Clinical Safety Recommendations

• Use as decision support only, never as final diagnosis.
• Always combine with human radiology review.
• Route uncertain/low-confidence outputs to manual review.
• Monitor sensitivity/specificity continuously on newly ingested data.

How to Use

Load base model + LoRA adapter

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel

base_model_id = "google/medgemma-1.5-4b-it"
adapter_id = "s0close/medgemma-dsa-lora"

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

Optional: merge adapter for standalone inference

Use: tools/publish_dsa_merged_model.py

Training Details

Training Data

• Internal DSA dataset prepared for vessel patency classification.
• Conversational VLM format (messages with image + text supervision).
• De-identified data expected; no PHI should be stored in the model artifacts.

Training Procedure

• QLoRA fine-tuning on Hugging Face Jobs.
• Quantization: 4-bit NF4 with double quantization.
• Compute dtype: bfloat16.

Hyperparameters (default project config)

• LoRA rank (r): 8
• LoRA alpha: 16
• LoRA dropout: 0.1
• Target modules: all-linear
• Learning rate: 2e-5 (script default)
• Batch size: 1
• Gradient accumulation: 8
• Epochs: 2 (script default)

Evaluation

Evaluation Setup

• Primary metrics: accuracy, sensitivity (blocked recall), specificity.
• Internal prior baseline showed a strong class-imbalance failure mode:
  • very high accuracy with near-zero blocked sensitivity when predicting majority class.

Current Results

• Run-specific results should be reported from the latest training_run.json and evaluation outputs in your HF run artifacts.
• Do not claim clinical-grade performance without external multi-site validation.

Technical Specifications

Architecture

• Base: MedGemma 1.5 4B instruction-tuned vision-language model.
• Adaptation: PEFT LoRA heads trained for DSA classification prompts.

Infrastructure

• Typical training flavor: Hugging Face Jobs on L4.
• Inference: CUDA/CPU depending on deployment constraints.

Ethical and Regulatory Notice

This model is a research artifact for hackathon/prototyping workflows. It is not a medical device and has not undergone regulatory approval. Any clinical use requires formal validation, governance, and compliance review.

Contact

For project-level questions, use your team repository issue tracker or designated team contact.