Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Google-MedGemma Based Abnormality Detection in Musculoskeletal radiographs

Soumyajit Maity, Pranjal Kamboj, Sneha Maity, Rajat Singh, Sankhadeep Chatterjee

TL;DR

Musculoskeletal radiograph abnormality detection is challenged by data diversity, subtle findings, and class imbalance. The paper proposes a MedGemma-based pipeline that leverages a SigLIP medical vision encoder to extract robust image embeddings, followed by a lightweight MLP classifier with selective unfreezing for domain adaptation. Across the MURA benchmark, the approach achieves high accuracy, macro-F1, and AUROC, outperforming conventional CNN- and autoencoder-based baselines and demonstrating efficient transfer learning. This work demonstrates the practicality of medical foundation models for scalable radiograph triage and outlines directions for multimodal integration, uncertainty quantification, and cross-institution validation.

Abstract

This paper proposes a MedGemma-based framework for automatic abnormality detection in musculoskeletal radiographs. Departing from conventional autoencoder and neural network pipelines, the proposed method leverages the MedGemma foundation model, incorporating a SigLIP-derived vision encoder pretrained on diverse medical imaging modalities. Preprocessed X-ray images are encoded into high-dimensional embeddings using the MedGemma vision backbone, which are subsequently passed through a lightweight multilayer perceptron for binary classification. Experimental assessment reveals that the MedGemma-driven classifier exhibits strong performance, exceeding conventional convolutional and autoencoder-based metrics. Additionally, the model leverages MedGemma's transfer learning capabilities, enhancing generalization and optimizing feature engineering. The integration of a modern medical foundation model not only enhances representation learning but also facilitates modular training strategies such as selective encoder block unfreezing for efficient domain adaptation. The findings suggest that MedGemma-powered classification systems can advance clinical radiograph triage by providing scalable and accurate abnormality detection, with potential for broader applications in automated medical image analysis. Keywords: Google MedGemma, MURA, Medical Image, Classification.

Google-MedGemma Based Abnormality Detection in Musculoskeletal radiographs

TL;DR

Musculoskeletal radiograph abnormality detection is challenged by data diversity, subtle findings, and class imbalance. The paper proposes a MedGemma-based pipeline that leverages a SigLIP medical vision encoder to extract robust image embeddings, followed by a lightweight MLP classifier with selective unfreezing for domain adaptation. Across the MURA benchmark, the approach achieves high accuracy, macro-F1, and AUROC, outperforming conventional CNN- and autoencoder-based baselines and demonstrating efficient transfer learning. This work demonstrates the practicality of medical foundation models for scalable radiograph triage and outlines directions for multimodal integration, uncertainty quantification, and cross-institution validation.

Abstract

This paper proposes a MedGemma-based framework for automatic abnormality detection in musculoskeletal radiographs. Departing from conventional autoencoder and neural network pipelines, the proposed method leverages the MedGemma foundation model, incorporating a SigLIP-derived vision encoder pretrained on diverse medical imaging modalities. Preprocessed X-ray images are encoded into high-dimensional embeddings using the MedGemma vision backbone, which are subsequently passed through a lightweight multilayer perceptron for binary classification. Experimental assessment reveals that the MedGemma-driven classifier exhibits strong performance, exceeding conventional convolutional and autoencoder-based metrics. Additionally, the model leverages MedGemma's transfer learning capabilities, enhancing generalization and optimizing feature engineering. The integration of a modern medical foundation model not only enhances representation learning but also facilitates modular training strategies such as selective encoder block unfreezing for efficient domain adaptation. The findings suggest that MedGemma-powered classification systems can advance clinical radiograph triage by providing scalable and accurate abnormality detection, with potential for broader applications in automated medical image analysis. Keywords: Google MedGemma, MURA, Medical Image, Classification.

Paper Structure

This paper contains 10 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Proposed Method
  4. Dataset
  5. Dataset Preprocessing
  6. Proposed Model Architecture
  7. Training Strategy
  8. Experimental Result
  9. Comparison with prior MURA experiments
  10. Conclusion

Figures (3)

  • Figure 1: Diagram of the proposed MedGemma based classifier.
  • Figure 2: Per-anatomy diagnostic performance of MedGemma-4B-PT across seven anatomical regions and overall.
  • Figure 3: Comparison of overall F1-score and AUROC for head-only fine-tuning and varying depth of encoder block unfreezing