MAD-UNet

MAD-UNet segments the pancreas in CT images using a U-shaped convolutional network enhanced with multiscale attention mechanisms, multiscale convolutions, and dense residual blocks to improve contextual and semantic feature extraction for accurate pancreas segmentation.

Key Features:

• Multiscale Attention Mechanism: Enriches contextual information to focus on relevant pancreatic regions and suppress unrelated areas, increasing sensitivity to marginal pancreatic details.
• Dense Residual Blocks: Facilitate deeper semantic feature learning and mitigate intraclass inconsistency via dense residual connections.
• Multiscale Convolutions: Capture features at multiple resolutions to represent both fine details and broader context.
• U-shaped Architecture: Employs an encoder–decoder U-Net style structure to preserve spatial resolution through skip connections for segmentation.
• Weighted Binary Cross-Entropy Loss: Applies class-weighted binary cross-entropy during optimization to address interclass indistinction.

Scientific Applications:

• Clinical pancreas segmentation: Provides precise segmentation of the pancreas from enhanced 3D CT scans, addressing high intrapatient variability and low contrast.
• Benchmark evaluation: Evaluated with fourfold cross-validation on the NIH-82 dataset (82 abdominal enhanced 3D CT scans) and the 2018 MICCAI segmentation decathlon (MSD, 281 3D CT scans), achieving mean Dice coefficients of 86.10% ± 3.52% on NIH-82 and 88.50% ± 3.70% on MSD.

Methodology:

U-shaped network architecture enhanced with multiscale attention mechanisms, multiscale convolutions, and dense residual blocks, trained using weighted binary cross-entropy loss to balance semantic and contextual feature learning.