Rubik's Cube+: A self-supervised feature learning framework for 3D medical image analysis.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Deep learning advances enable automated 3D medical data analysis.
• Manual annotation of 3D medical data is time-consuming and labor-intensive, hindering neural network training.
• Self-supervised learning offers a solution by reducing reliance on extensive annotations.

Purpose of the Study:

• To propose a novel self-supervised learning framework for 3D volumetric medical data.
• To address the challenge of limited annotated data for training deep learning models in medical imaging.

Main Methods:

• Introduced a new self-supervised learning framework for 3D medical data.
• Developed a pretext task named Rubik's cube+ for pre-training 3D neural networks.
• The Rubik's cube+ task includes cube ordering, rotation, and masking to learn invariant features and tolerate noise.

Main Results:

• Pre-training 3D neural networks with Rubik's cube+ significantly improved accuracy on downstream tasks.
• Fine-tuning from Rubik's cube+ pre-trained weights boosted performance in cerebral hemorrhage classification.
• Enhanced performance was also observed in brain tumor segmentation tasks.

Conclusions:

• The proposed Rubik's cube+ self-supervised learning framework effectively pre-trains 3D neural networks.
• This approach alleviates the need for large annotated datasets in medical imaging.
• It offers a promising direction for improving automated analysis of 3D medical data.