Automatic medical imaging segmentation via self-supervising large-scale convolutional neural networks.

Area of Science:

• Deep learning for medical image analysis
• Self-supervised learning in radiology
• Computational anatomy and image segmentation

Background:

• Supervised learning for medical image segmentation requires extensive labeled data, which is costly and time-consuming to acquire.
• Clinical data variability poses challenges for developing generalizable segmentation models.
• Self-supervised learning offers a promising avenue to leverage large unlabeled datasets.

Purpose of the Study:

• To develop a robust, large-scale deep learning model for medical image segmentation.
• To utilize self-supervised learning to overcome limitations of supervised methods and data variability.
• To create a scalable model applicable across various imaging modalities.

Main Methods:

• Curated a multi-center CT dataset for self-supervised pre-training using masked image modeling and sparse submanifold convolution.
• Designed and pre-trained Sparse Submanifold U-Nets (SS-UNets) of varying sizes.
• Fine-tuned SS-UNets on the TotalSegmentator dataset and evaluated robustness and transferability on unseen datasets.

Main Results:

• SS-UNets outperformed state-of-the-art self-supervised methods, achieving high Dice Similarity Coefficient (DSC) and Surface Dice Coefficient (SDC) scores.
• SS-UNet-B achieved 84.3% DSC and 88.0% SDC on TotalSegmentator.
• Segmentation performance scaled with model size, showing significant improvements with larger parameter counts.

Conclusions:

• Self-supervised learning is effective for medical image segmentation across CT, MRI, and PET domains.
• The developed approach reduces reliance on labeled data, mitigates overfitting, and enhances generalizability.
• Potential applications include streamlined cancer detection and radiotherapy planning through accurate organ and lesion segmentation.