MsVRL: Self-Supervised Multiscale Visual Representation Learning via Cross-Level Consistency for Medical Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated medical image segmentation is crucial for diagnosis and treatment planning.
• Training accurate models is challenging due to extensive annotation requirements, especially for 3D images.
• Self-supervised learning offers a promising avenue for reducing annotation burden in medical imaging.

Purpose of the Study:

• To develop a novel self-supervised learning model for medical image segmentation.
• To address limitations of global representation learning in non-iconic medical imaging tasks with varying scales.
• To improve the accuracy and robustness of automated medical image segmentation.

Main Methods:

• Proposed the Multi-scale Visual Representation self-supervised Learning (MsVRL) model.
• Incorporated multi-scale representation, canvas matching, embedding pre-sampling, a center-ness branch, and cross-level consistent loss.
• Pre-trained MsVRL on unlabeled datasets (RibFrac, MSD) and evaluated on downstream segmentation tasks (BCV, MSD spleen, KiTS).

Main Results:

• MsVRL demonstrated superior performance in medical image segmentation tasks.
• The model effectively handles different target scales and performs finer-grained representation learning.
• Achieved state-of-the-art results compared to existing methods on tested datasets.

Conclusions:

• The proposed MsVRL model significantly advances self-supervised learning for medical image segmentation.
• MsVRL effectively alleviates the need for extensive manual annotations.
• This approach holds great potential for improving clinical diagnosis and treatment planning through enhanced automated segmentation.