MedImageInsight for Thoracic Cavity Health Classification from Chest X-rays

TL;DR

This work evaluates MedImageInsight for binary chest X-ray classification (Normal vs Abnormal) using two paths: end-to-end fine-tuning and transfer-learning with embeddings, on ChestX-ray14 plus real-world clinical data. The fine-tuned model achieves a ROC-AUC of $0.888$ with superior calibration, rivaling established methods like CheXNet, and supports scalable integration into PACS and web-based triage workflows. Transfer-learning with embeddings lags behind in both discrimination and calibration, though SVM offers its best performance among the baselines. The findings support the potential of healthcare foundation models to reduce radiologist workload and enable rapid, reliable triage, with future work extending to multi-label pathology and prospective validations.

Abstract

Chest radiography remains one of the most widely used imaging modalities for thoracic diagnosis, yet increasing imaging volumes and radiologist workload continue to challenge timely interpretation. In this work, we investigate the use of MedImageInsight, a medical imaging foundational model, for automated binary classification of chest X-rays into Normal and Abnormal categories. Two approaches were evaluated: (1) fine-tuning MedImageInsight for end-to-end classification, and (2) employing the model as a feature extractor for a transfer learning pipeline using traditional machine learning classifiers. Experiments were conducted using a combination of the ChestX-ray14 dataset and real-world clinical data sourced from partner hospitals. The fine-tuned classifier achieved the highest performance, with an ROC-AUC of 0.888 and superior calibration compared to the transfer learning models, demonstrating performance comparable to established architectures such as CheXNet. These results highlight the effectiveness of foundational medical imaging models in reducing task-specific training requirements while maintaining diagnostic reliability. The system is designed for integration into web-based and hospital PACS workflows to support triage and reduce radiologist burden. Future work will extend the model to multi-label pathology classification to provide preliminary diagnostic interpretation in clinical environments.

Figures (5)

• Figure 1: Receiver Operating Characteristic (ROC) curves for the evaluated classifiers (SVM, Random Forest, MLP, Logistic Regression, and KNN) in the transfer learning approach.
• Figure 2: Precision-Recall curves of the finetuned classifier (a) (on the left) and other classifiers of the transfer learning approach (b) (on the right).
• Figure 3: Calibration plots of the finetuned classifier (a)(on the left) and other classifiers of the transfer learning approach (b) (on the right).
• Figure 4: Comparison of ROC curves for three models: (a) (on top left) Our model, (b) (on top right) Wong et al. [14], and (c) (at bottom) CheXNet by Rajpurkar et al. [3]. All curves show the trade-off between sensitivity (true positive rate) and 1 specificity (false positive rate).
• Figure 5: Workflow comparison of two model training approaches for chest X-ray image classification. (a) (on the left) End-to-end fine-tuning pipeline using the MedImageInsight model as a base network, where images undergo preprocessing before fine-tuning with labelled data (Normal vs. Abnormal). (b) (on the right) Feature extraction approach where embeddings generated by MedImageInsight are input to traditional machine learning classifiers (KNN, SVM, Random Forest, Logistic Regression, and MLP) to predict Normal vs. Abnormal.