Volume Fusion-Based Self-Supervised Pretraining for 3D Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning for medical image segmentation struggles with limited annotated data.
• Self-Supervised Learning (SSL) offers a solution by leveraging unannotated images for feature learning.
• Current SSL methods have limitations in pretraining decoders for downstream segmentation tasks.

Purpose of the Study:

• To introduce Volume Fusion (VolF), a novel SSL strategy for pretraining 3D medical image segmentation models.
• To minimize the gap between pretext and downstream tasks in SSL for improved segmentation.
• To enable effective training of segmentation models using unannotated data.

Main Methods:

• Proposed Volume Fusion (VolF) strategy for pretraining 3D segmentation models.
• Introduced a pseudo-segmentation pretext task involving fused sub-volumes and a fusion coefficient map.
• Utilized standard supervised segmentation loss functions for training without manual annotations.

Main Results:

• VolF demonstrated significant performance gains compared to training from scratch.
• Achieved faster convergence speeds in pretraining 3D segmentation models.
• Outperformed several state-of-the-art SSL methods on abdominal CT datasets and generalized to different body parts and modalities.

Conclusions:

• Volume Fusion (VolF) is an effective SSL strategy for pretraining 3D medical image segmentation models.
• VolF offers improved performance, faster convergence, and better generalizability than existing SSL methods.
• The proposed method is versatile and applicable to various network architectures and imaging modalities.