Self-supervised learning for chest computed tomography: training strategies and effect on downstream applications

Self-supervised pre-training can reduce the amount of labeled training data needed by pre-learning fundamental visual characteristics of the medical imaging data. We investigate several self-supervised training strategies for chest computed tomography exams and their effects on downstream applications.

We benchmark five well-known self-supervision strategies (masked image region prediction, next slice prediction, rotation prediction, flip prediction, and denoising) on 15 M chest computed tomography (CT) slices collected from four sites of the Mayo Clinic enterprise, United States. These models were evaluated for two downstream tasks on public datasets: pulmonary embolism (PE) detection (classification) and lung nodule segmentation. Image embeddings generated by these models were also evaluated for prediction of patient age, race, and gender to study inherent biases in models' understanding of chest CT exams.

The use of pre-training weights especially masked region prediction-based weights, improved performance, and reduced computational effort needed for downstream tasks compared with task-specific state-of-the-art (SOTA) models. Performance improvement for PE detection was observed for training dataset sizes as large as $\sim 380K$ with a maximum gain of 5% over SOTA. The segmentation model initialized with pre-training weights learned twice as fast as the randomly initialized model. While gender and age predictors built using self-supervised training weights showed no performance improvement over randomly initialized predictors, the race predictor experienced a 10% performance boost when using self-supervised training weights.

We released self-supervised models and weights under an open-source academic license. These models can then be fine-tuned with limited task-specific annotated data for a variety of downstream imaging tasks, thus accelerating research in biomedical imaging informatics.