Parts2Whole: Self-supervised Contrastive Learning via Reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Contrastive representation learning excels in computer vision but has limitations for 3D medical imaging due to high computational demands.
• Reconstruction-based self-supervised learning is effective for 3D medical imaging but lacks contrastive representation learning capabilities.

Purpose of the Study:

• To introduce Parts2Whole, a novel framework for self-supervised contrastive learning via reconstruction tailored for 3D medical imaging.
• To learn effective contrastive representations by exploiting the intrinsic part-whole relationship without relying on contrastive loss.

Main Methods:

• The Parts2Whole framework reconstructs images (wholes) from their constituent parts.
• This reconstruction process implicitly enforces learning similar latent features for parts of the same whole and dissimilar features for parts of different wholes.
• The method leverages the universal and intrinsic part-whole relationship for representation learning.

Main Results:

• Parts2Whole was evaluated on five diverse 3D medical imaging tasks, including classification and segmentation.
• The framework demonstrated superior performance on two tasks and competitive results on the remaining three when compared to four established 3D pre-trained models.
• The outperformance is attributed to the effective contrastive representations learned by the Parts2Whole framework.

Conclusions:

• Parts2Whole offers a viable and effective approach for self-supervised contrastive learning in 3D medical imaging.
• The framework successfully learns discriminative representations by utilizing reconstruction from part-whole relationships.
• The study provides valuable insights into improving self-supervised learning methodologies for medical image analysis.