Diversity-driven MG-MAE: Multi-granularity representation learning for non-salient object segmentation.

Area of Science:

• Artificial Intelligence
• Medical Image Analysis
• Computer Vision

Background:

• Masked Autoencoders (MAEs) are effective self-supervised learning models for image analysis.
• MAEs struggle with feature diversity for non-salient medical structures due to dimensional collapse.
• Accurate segmentation of non-salient objects is crucial in medical imaging.

Purpose of the Study:

• To propose a Multi-Granularity Masked Autoencoder (MG-MAE) framework to enhance feature diversity for non-salient object segmentation.
• To address the dimensional collapse problem in MAEs for medical image analysis.
• To improve the discrimination of fine-grained patterns in medical images.

Main Methods:

• Developed a multi-granularity framework with global and local branches for hierarchical feature representation.
• Incorporated a diversity-enhanced loss function with Nuclear Norm Maximization (NNM) to prevent feature space collapse.
• Implemented a Dynamic Weight Adjustment (DWA) strategy to focus on challenging regions using entropy-driven modulation.

Main Results:

• MG-MAE demonstrated statistically significant improvements in Dice Similarity Coefficient (DSC) scores across five clinical datasets.
• The framework successfully improved segmentation of non-salient objects compared to state-of-the-art methods.
• Achieved enhanced feature diversity crucial for discriminating subtle anatomical structures and pathologies.

Conclusions:

• MG-MAE effectively overcomes the limitations of conventional MAEs in medical image segmentation.
• The proposed framework offers a robust solution for segmenting non-salient structures in medical imaging.
• MG-MAE represents a significant advancement in self-supervised learning for medical applications.