MSU-Net

MSU-Net leverages multi-scale convolutional sequences and diverse receptive fields to improve medical image segmentation by extracting richer semantic features.

Key Features:

• Multi-Scale Convolutional Sequences: Employs multiple convolution sequences to extract richer semantic features from medical images for segmentation.
• Diverse Receptive Fields: Integrates convolution kernels with varying receptive fields to capture a broader range of spatial information.
• Adaptive Network Width: Addresses unknown optimal network width by combining convolution kernels with different receptive fields to adjust network capacity for segmentation tasks.
• Extended Multi-Scale Blocks: Extends the multi-scale block design to other U-Net variants to apply the approach across different network configurations.

Scientific Applications:

• Cross-modality medical image segmentation: Evaluated on five medical image segmentation datasets covering electron microscopy, dermoscopy, and ultrasound.
• Quantitative evaluation with Intersection over Union (IoU): Reported IoU scores of 0.771, 0.867, 0.708, 0.900, and 0.702 on the five datasets.

Methodology:

Combines multi-scale convolutional sequences and diverse receptive fields to enhance feature extraction and network adaptability, addressing limitations of fixed receptive fields and unknown optimal network widths.