MSCF-Net: A Vision Mamba Network with Multi-Scale Context Bridging and Cross-Layer Adaptive Fusion for Medical Image...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Accurate medical image segmentation is difficult due to lesion variations and background interference.
• Existing Mamba-based U-Net models struggle with local multi-scale representation and skip connections.

Purpose of the Study:

• To propose MSCF-Net, a novel Vision Mamba segmentation network for improved dermoscopic and endoscopic image analysis.
• To address limitations in local context and feature fusion in Mamba-based segmentation networks.

Main Methods:

• Developed MSCF-Net, a Vision Mamba network built upon VM-UNet.
• Introduced Multi-Scale Context Bridging (MSCB) to enrich bottleneck features with diverse context.
• Implemented Cross-Layer Adaptive Fusion (CLAF) for adaptive recalibration of encoder-decoder features.
• Utilized a structure loss to enhance region completeness and boundary quality.

Main Results:

• MSCF-Net achieved high performance on benchmark datasets (ISIC 2017, ISIC 2018, CVC-ClinicDB) with Dice scores up to 91.72% and mIoU up to 84.56%.
• Demonstrated competitive segmentation accuracy compared to existing methods.
• Ablation studies confirmed MSCB's role in scale-aware representation and CLAF's effectiveness in focusing on lesion cues.

Conclusions:

• MSCF-Net offers a favorable accuracy-efficiency trade-off for medical image segmentation.
• The proposed MSCB and CLAF modules significantly improve segmentation quality.
• MSCF-Net shows promise for clinical applications requiring precise medical image segmentation.