Skip to content

MedImageInsight for Thoracic Cavity Health Classification from Chest X-rays

🕒 Published (v1): 2025-11-21 08:42 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper evaluates Microsoft's MedImageInsight foundation model for binary Normal/Abnormal chest X-ray classification via two paradigms: end-to-end fine-tuning and transfer learning with frozen embeddings fed to traditional ML classifiers. The fine-tuned variant achieves ROC-AUC 0.888, matching CheXNet-level performance while requiring less task-specific data. The system targets integration into PACS and web-based hospital triage workflows.

Problem

Radiologist workload has grown dramatically over the past 15 years, creating pressure for automated triage tools. Existing CNN-based classifiers are narrow and data-hungry; the question is whether a general-purpose medical imaging foundation model (MedImageInsight) can be efficiently adapted for chest X-ray triage without the resource overhead of training from scratch.

Method

Two adaptation strategies were compared for binary (Normal vs. Abnormal) chest X-ray classification using MedImageInsight:

  1. Fine-tuning: End-to-end adaptation of MedImageInsight on a combined dataset (ChestX-ray14 + real-world data from Indian and UK partner hospitals). Training used AdamW (lr=1e-5, weight decay=0.2), batch size 48 on V100 GPUs, cosine decay with warmup.
  2. Transfer learning: MedImageInsight used as a frozen feature extractor; embeddings fed to five downstream classifiers (KNN, Logistic Regression, SVM, Random Forest, MLP) optimized via grid search with 5-fold cross-validation on a 10,000-image stratified sample from ChestX-ray14.

All 14 ChestX-ray14 pathology labels were merged into a single "Abnormal" class; "No Finding" became "Normal." Evaluation metrics included ROC-AUC, F1, Brier Score (calibration), accuracy, precision, recall, sensitivity, and specificity on a held-out test set.

Key Contributions

  • Systematic comparison of fine-tuning vs. embedding-based transfer learning for a medical imaging foundation model on chest X-ray binary triage.
  • Evaluation on a combined public (ChestX-ray14) and real-world multi-institution clinical dataset (India + UK hospitals), not solely a public benchmark.
  • Demonstration that fine-tuned MedImageInsight achieves calibration (Brier Score 0.137) superior to all transfer learning baselines, relevant for clinical decision support.
  • System design targeting direct PACS and web-based integration for hospital triage.

Results

  • Fine-tuned classifier: ROC-AUC 0.888, Accuracy 0.786, Precision 0.875, Recall 0.723, F1 0.792, Brier Score 0.137 — best across all metrics.
  • Best transfer learning baseline (SVM): ROC-AUC 0.828, Accuracy 0.762, Precision 0.724, Recall 0.740, F1 0.732, Brier Score 0.168.
  • Fine-tuned model ROC-AUC (0.888) is reported as comparable to CheXNet (Rajpurkar et al.) and Wong et al. architectures based on published ROC curve comparisons.
  • Among transfer learning classifiers: SVM > LogReg > RF > MLP > KNN on ROC-AUC (0.828, 0.820, 0.818, 0.804, 0.760 respectively).

Limitations

  • Binary classification only (Normal vs. Abnormal); no per-pathology multi-label output.
  • Retrospective evaluation only; no prospective validation or radiologist-in-the-loop assessment.
  • Transfer learning experiments used only 10,000 images sampled from the combined dataset, not the full training set.
  • No ablation on the contribution of real-world hospital data vs. ChestX-ray14 alone.
  • Brier Scores and other metrics are reported on a combined test set whose exact size and institution split are not detailed.
  • Fine-tuning dataset composition (hospital split, total N) is shown only as class-label counts, not total image counts.

Relevance to Foundation Models in Medicine

This paper provides a direct applied benchmark of MedImageInsight, Microsoft's open-source medical imaging foundation model, in a clinically realistic setting that mixes public and proprietary hospital data — a test condition rarely covered in foundation model papers. The finding that fine-tuning outperforms frozen-embedding transfer learning on both discriminative and calibration metrics informs the broader question of how much task-specific adaptation medical foundation models require. The PACS integration focus positions this as an early deployment case study for foundation-model-powered radiology triage, relevant to researchers tracking how general-purpose medical vision encoders translate into clinical workflows.