ReFs: A hybrid pre-training paradigm for 3D medical image segmentation

Area of Science:

• Artificial Intelligence
• Computer Vision
• Medical Imaging

Background:

• Self-supervised learning (SSL) excels in medical image segmentation but often uses a two-stage process where SSL is unaware of downstream tasks, leading to suboptimal feature representations.
• The standard SSL paradigm involves unsupervised representation learning followed by supervised fine-tuning, which can limit performance due to task-agnostic pre-training.

Purpose of the Study:

• To introduce a hybrid pre-training paradigm, reference-guided self-supervised learning (ReFs), that integrates supervised objectives into the SSL process.
• To enhance feature representation quality for medical image segmentation by making SSL aware of the downstream task.

Main Methods:

• Proposed a hybrid pre-training paradigm combining self-supervised and supervised objectives.
• Incorporated an off-the-shelf medical image segmentation task as a supervised reference during SSL.
• Developed a gradient matching method to align feature extractor updates from both SSL and reference tasks.

Main Results:

• Demonstrated improved representation quality by encouraging low prediction loss on both SSL and reference tasks.
• Showcased effectiveness on seven downstream medical image segmentation benchmarks.
• Validated the ReFs paradigm on large-scale unlabeled and reference datasets.

Conclusions:

• The proposed reference-guided self-supervised learning (ReFs) paradigm effectively improves medical image segmentation.
• Integrating supervised tasks into SSL enhances feature representations by aligning learning objectives.
• ReFs offers a promising direction for advancing self-supervised learning in medical image analysis.