Medical Transformer: Universal Encoder for 3-D Brain MRI Analysis.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Transfer Learning

Background:

• Limited annotated 3D medical datasets hinder deep learning model training.
• Transfer learning is crucial for data-driven medical image analysis.

Purpose of the Study:

• Propose Medical Transformer, a novel transfer learning framework for 3D medical images.
• Improve high-level representation in 3D volumes by modeling them as sequences of 2D slices.

Main Methods:

• Utilize a multiview approach leveraging information from three planes of 3D volumes.
• Pretrain the model using self-supervised learning (SSL) for masked encoding vector prediction on a large-scale healthy brain MRI dataset.
• Evaluate the pretrained model on brain disease diagnosis, brain age prediction, and brain tumor segmentation.

Main Results:

• Medical Transformer outperforms state-of-the-art transfer learning methods.
• Achieved significant parameter reduction: up to 92% for classification/regression, 97% for segmentation.
• Demonstrated good performance even with partial training samples.

Conclusions:

• Medical Transformer offers an efficient and effective transfer learning solution for 3D medical image analysis.
• The framework shows promise for various brain MRI research tasks, especially with limited data.