Neural Memory State Space Models for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Deep learning, particularly UNet, is vital for medical image segmentation.
• Transformers improve UNet but face computational challenges.
• State-Space Models (SSMs) like Mamba offer linear complexity, and neural memory Ordinary Differential Equations (nmODEs) show promise.

Purpose of the Study:

• To explore the strengths and weaknesses of nmODEs and SSMs.
• To propose a novel nmSSM decoder architecture combining their advantages.
• To validate the effectiveness of nmSSM-UNet for medical image segmentation.

Main Methods:

• Developed a novel nmSSM decoder integrating nmODEs and SSMs.
• Constructed the nmSSM-UNet architecture.
• Conducted experiments on PH2, ISIC2018, and BU-COCO datasets.
• Performed ablation studies to confirm improvements.

Main Results:

• nmSSM-UNet demonstrates powerful nonlinear representation and global information processing.
• The architecture shows promising results in medical image segmentation tasks.
• Ablation experiments validated the contributions of the proposed improvements.

Conclusions:

• The nmSSM decoder effectively combines the strengths of nmODEs and SSMs.
• nmSSM-UNet is a valuable tool for advancing medical image segmentation.
• This work highlights the potential of integrating nmODEs and SSMs for improved medical AI.