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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Self-supervised pre-training (SSP) leverages unlabeled data to learn fundamental visual features in medical imaging.
• This reduces the dependency on large, manually annotated datasets, which are costly and time-consuming to acquire.
• Chest computed tomography (CT) exams are a critical modality, but analyzing them requires significant expertise.

Purpose of the Study:

• To investigate the efficacy of various self-supervised pre-training strategies for chest CT imaging.
• To evaluate the impact of these strategies on downstream tasks such as pulmonary embolism detection and lung nodule segmentation.
• To assess potential biases in self-supervised models by evaluating their performance in predicting patient demographics.

Main Methods:

• Benchmarked five self-supervision strategies: masked image region prediction, next slice prediction, rotation prediction, flip prediction, and denoising.
• Utilized a large dataset of 15 million chest CT slices from four Mayo Clinic sites.
• Evaluated model performance on public datasets for pulmonary embolism detection and lung nodule segmentation, and for predicting patient age, race, and gender.

Main Results:

• Pre-training with masked region prediction significantly improved performance and reduced computational effort compared to state-of-the-art models.
• A 5% performance gain was observed for pulmonary embolism detection, even with limited training data.
• Segmentation models initialized with pre-trained weights learned twice as fast as randomly initialized models.
• A 10% performance boost was noted for predicting patient race using self-supervised weights, while age and gender prediction showed no improvement.

Conclusions:

• Self-supervised pre-training, particularly masked region prediction, offers substantial benefits for chest CT analysis.
• These methods accelerate research by enabling fine-tuning with limited annotated data for various downstream tasks.
• The study released open-source models and weights to facilitate further research in biomedical imaging informatics.