General-Purpose Chest X-ray Screening using AI Foundation Model

Introduction: Chest X-ray (CXR) is the most commonly ordered imaging test worldwide. Clinical practice of CXR examination involves a wide range of tasks, from disease detection of the lungs, heart, blood vessel, and bone to monitoring postoperative recovery and support device positions. Training specialist artificial intelligence (AI) models for each task from scratch has fundamental weaknesses. Models trained with small-scale, monotonous data will have significant performance discrepancies between development and deployment due to domain shifts. Self-supervised learning has been invented to leverage abundant, low-cost unlabeled data to learn a generalist model useful for a wide range of tasks. However, the existing self-supervised AI models for CXR analysis remain constrained by the limited pretraining data size and diversity, given that the widely-used CheXpert[1], MIMIC-CXR[2], or NIH Chest X-ray14[3] datasets are all curated from single centers containing around half a million images in total. Such limits in data size and diversity underrate the scalability of the modern AI deep learning models. In this work, we posit that foundation models pretrained with self-supervised learning and large-scale, diverse CXR images can attain improved generalization performance in a variety of CXR disease detection tasks.

Materials and

Methods: To curate a large-scale, diverse dataset for CXR foundation model pretraining, we collected images from multiple public CXR datasets, including NIH Chest X-ray14, MIMIC-CXR, CheXpert, BRAX[4], and PadChest[5], totaling 875,732 CXR images as shown in Table 1. The training and validation sets of NIH Chest X-ray14, MIMIC-CXR, CheXpert, and BRAX were included for pretraining, while their test set was kept unseen. To enable the model to learn meaningful representations from unlabeled CXR images, we adopted the state-of-the-art self-supervised learning method DINOv2[6] and the ViT-L transformer network architecture[7]. The pretrained model was evaluated via linear probing, where a learnable linear classifier was incorporated to process the output features from the frozen pretrained model to produce predictions. The linear probe classifiers were trained and tested on 191,229 CXRs and 518 CXRs respectively for thoracic disease detection in CheXpert, and on 528 CXRs and 134 CXRs for tuberculosis detection in Shenzhen.

Results, Conclusions, and Discussions: Results
We tested our pretrained CXR foundation model (CXR-FM) on CheXpert and Shenzhen tuberculosis detection[8] datasets via linear probing. Experimental results in Table 2 show that CXR-FM achieved an average AUROC of 0.874 on the detection of five CheXpert competition pathologies (cardiomegaly, edema, consolidation, atelectasis, and pleural effusion), outperforming the previous method MoCo-CXR pretrained merely on MIMIC-CXR[2] by a large margin (+0.077). Table 3 shows that CXR-FM achieved comparable performance with MoCo-CXR[9] on the Shenzhen tuberculosis detection task.

Conclusions
This study curated a large-scale, diverse dataset containing more than 800K CXR images from multiple centers for CXR foundation model pretraining. The self-supervised model pretrained on this dataset outperform the previous method pretrained on a single-center dataset, demonstrating the effectiveness of self-supervised pretraining with large-scale, diverse data.

Discussions
This study shows the promising performance of large-scale self-supervised pretraining to develop foundation models. Abundant unlabeled CXR images available to retrieve from public sources motivates us to curate an even larger-scale CXR dataset for more effective self-supervised pretraining in the future. Besides the common thoracic disease screening tasks in this study, extended computer-aided CXR diagnosis tasks such as cardiovascular disease risk prediction and mortality prognostication are of our interest to further evaluate the generalization performance.

Acknowledgements (Optional):