H2M-UNet: Hierarchical Memory Mamba-Driven UNet Collaborative Optimization Based on Long-Range Forgetting Mitigation...

Area of Science:

• Medical Image Analysis
• Deep Learning Architectures
• Computer Vision

Background:

• Conventional Mamba networks struggle with long-range dependencies in medical image segmentation due to local scanning, leading to prediction inconsistencies.
• This limitation compromises segmentation accuracy, particularly for long image sequences.

Purpose of the Study:

• To propose a novel Hierarchical Memory Mamba-UNet (H2M-UNet) to overcome the long-range forgetting issue in Mamba-based medical image segmentation.
• To enhance long-range dependency modeling and improve prediction consistency and segmentation accuracy.

Main Methods:

• Introduction of a Hierarchical Memory 2D (HM2D) scanning mechanism with a four-branch parallel strategy for multi-scale feature co-optimization.
• HM2D utilizes global, local refinement, and conventional dual-path scanning for differentiated feature processing at various resolutions.
• Integration of HM2D into the UNet architecture to create the H2M-UNet model.

Main Results:

• H2M-UNet demonstrated superior performance on Synapse (multi-organ) and ACDC (cardiac) segmentation datasets.
• Achieved an average Dice score of 81.79% and HD95 of 14.00 on Synapse, outperforming existing methods.
• Attained an average Dice of 91.33% and HD95 of 1.12 on ACDC, showing significant improvements.

Conclusions:

• The HM2D mechanism effectively mitigates long-range forgetting and preserves fine-grained features in medical image segmentation.
• H2M-UNet achieves superior segmentation accuracy and boundary consistency in complex medical imaging scenarios.
• The proposed method offers a promising advancement for deep learning-based medical image segmentation.