Self-supervised 3D anatomy segmentation using self-distilled masked image transformer (SMIT).

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Vision transformers excel at long-range context modeling for image segmentation but require extensive labeled data.
• Self-supervised learning (SSL) has shown promise in medical image segmentation, particularly with convolutional networks.
• Acquiring large labeled datasets for medical imaging remains a significant challenge.

Purpose of the Study:

• To develop a novel self-supervised learning method for vision transformers (SMIT) tailored for 3D multi-organ segmentation.
• To address the data scarcity issue in medical image analysis by leveraging self-distillation and masked image modeling.
• To improve the accuracy and reduce data requirements for medical image segmentation tasks.

Main Methods:

• Developed a self-distillation learning with masked image modeling approach (SMIT) for pre-training vision transformers.
• Employed a dense pixel-wise regression pretext task (masked image prediction) within masked patches.
• Incorporated masked patch token distillation for effective pre-training.
• Utilized a diverse dataset of 3,643 CT scans from various cancer types and COVID-19 for pre-training.

Main Results:

• Achieved high accuracy in 3D multi-organ segmentation on both MRI (average DSC of 0.875) and CT (average DSC of 0.878) datasets.
• Demonstrated superior performance compared to other pretext tasks, requiring significantly fewer fine-tuning datasets.
• Showcased clear accuracy improvements over existing SSL methods in extensive comparisons.

Conclusions:

• SMIT effectively enables self-supervised pre-training of vision transformers for medical image segmentation.
• The proposed method significantly reduces the need for large labeled datasets in medical imaging tasks.
• SMIT offers a promising solution for accurate and data-efficient 3D multi-organ segmentation.