CA-Net

CA-Net enhances medical image segmentation by integrating joint spatial, channel, and scale attention mechanisms within a convolutional neural network to improve segmentation accuracy and explainability.

Key Features:

• Comprehensive Attention Mechanisms: Incorporates Joint Spatial Attention, Channel Attention, and Scale Attention modules to focus on foreground regions, recalibrate channel-wise feature responses, and emphasize salient multi-scale feature maps respectively.
• Explainability and Visualization: Visualizes attention weight maps to expose the network's focus areas and support interpretation of segmentation decisions.
• Performance and Efficiency: Validated on ISIC 2018 and multi-class fetal MRI datasets with Dice improvements for skin lesion segmentation (87.77% to 92.08%), placenta segmentation (84.79% to 87.08%), and fetal brain segmentation (93.20% to 95.88%), while using a model size approximately 15 times smaller than DeepLabv3+.

Scientific Applications:

• Medical image segmentation: Improves automated delineation of anatomical structures across varied imaging scenarios where accurate segmentation is required for analysis or intervention planning.
• Dermatology (skin lesion analysis): Enhances segmentation of skin lesions on datasets such as ISIC 2018 to support lesion characterization and downstream analysis.
• Obstetrics (fetal MRI): Supports multi-class fetal MRI segmentation tasks including placenta and fetal brain delineation.

Methodology:

Integrates joint spatial, channel, and scale attention modules within a convolutional neural network architecture to increase sensitivity to spatial, channel, and scale variations in medical images.