Clinical target segmentation using a novel deep neural network: double attention Res-U-Net

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate medical image segmentation is crucial but challenged by noise, signal dropout, and complex object modeling.
• Existing segmentation methods struggle with diverse medical imaging modalities and intricate targets.

Purpose of the Study:

• To introduce a novel Double Attention Res-U-Net architecture for improved medical image segmentation.
• To address limitations of baseline segmentation approaches in complex medical imaging scenarios.

Main Methods:

• Developed a U-Net-based model with two consecutive networks (five and four encoding/decoding levels).
• Incorporated residual blocks and skip connections to mitigate vanishing gradient problems.
• Utilized multi-scale attention gates for richer contextual information extraction.

Main Results:

• Achieved high Dice and Jaccard coefficients: 95.79% and 91.62% for CRL segmentation, 93.84% and 89.08% for fetal foot segmentation.
• Outperformed state-of-the-art U-Net models on CVC-ClinicDB (83% Dice, 75.31% Jaccard) and multi-site MRI datasets (92.07% Dice, 87.14% Jaccard).

Conclusions:

• The Double Attention Res-U-Net architecture demonstrates significant improvements in medical image segmentation accuracy.
• The proposed model effectively handles challenges across different medical imaging systems and datasets.