Tailored self-supervised pretraining improves brain MRI diagnostic models.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Informatics

Background:

• Self-supervised learning (SSL) shows promise for deep learning but is underutilized in brain MRI analysis.
• Large-scale, unlabeled brain MRI datasets are valuable for improving AI models in clinical settings.

Purpose of the Study:

• To leverage large, public brain MRI datasets for self-supervised pretraining of deep learning models.
• To enhance performance in downstream clinical tasks, aiding clinical decision support systems.

Main Methods:

• Developed data filtering techniques (image entropy, slice position) to curate a focused dataset of 250K brain MRI images.
• Applied Momentum Contrast (MoCo) v3 for self-supervised feature learning on the curated brain MRI dataset.
• Evaluated pretrained models on tumor classification, lesion detection, hippocampal segmentation, and MRI reconstruction.

Main Results:

• Brain MRI-specific pretraining surpassed ImageNet and general medical dataset pretraining.
• Achieved a ~2.8% increase in 4-class tumor classification accuracy.
• Improved tumor detection mean average precision by ~0.9%, hippocampal segmentation Dice score by ~3.6%, and reconstruction PSNR by ~0.1.

Conclusions:

• Self-supervised learning on large, tailored brain MRI datasets is effective for enhancing AI model performance.
• This approach shows significant potential for advancing clinical decision support systems in neuroimaging.
• Publicly available data can be effectively utilized to create specialized datasets for medical AI pretraining.