FAM

FAM enhances lesion segmentation in computed tomography (CT) images for COVID-19 pneumonia by applying a dual attention mechanism to improve spatial localization and channel-wise feature weighting.

Key Features:

• Dual Attention Mechanism: Incorporates both a channel attention module to emphasize important feature maps and a spatial attention module to focus on significant image regions.
• Spatial Attention Module: Generates a rough spatial attention map using median filtering and distance transformation as a lesion shape prior and refines it via two 7×7 convolutional layers.
• Integration with Segmentation Networks: Has been integrated with six state-of-the-art segmentation networks, including UNet and DeepLabV3+.
• Training Efficiency: Accelerates model convergence during training, improving training speed for deep learning segmentation models.

Scientific Applications:

• Improved Segmentation Accuracy: Increases Dice Similarity Coefficient (DSC) by approximately 2% for lesion segmentation tasks.
• Reduction in False Negatives and Positives: Reduces false negatives and false positives by up to 17.6%, outperforming attention mechanisms such as CBAM and SENet.
• COVID-19 Lesion Segmentation: Applied to automatic segmentation of COVID-19 pneumonia lesions from CT images to improve lesion localization and quantification.

Methodology:

Generate a preliminary spatial attention map from CT images using median filtering and distance transformation to form a shape prior, refine this map with two 7×7 convolutional layers, and combine the refined spatial attention with a channel attention module.