FANet

FANet improves biomedical image segmentation by leveraging feedback attention across training epochs to iteratively refine segmentation predictions.

Key Features:

• Epoch Information Utilization: FANet prunes prediction maps from each training epoch so information from earlier epochs contributes to subsequent training and predictions.
• Novel Architecture: The architecture integrates the previous epoch's mask with the current epoch's feature map to apply hard attention across convolutional layers.
• Iterative Prediction Rectification: FANet performs iterative rectification of predictions at test time, refining segmentation outputs through feedback-based iterations.

Scientific Applications:

• Biomedical image segmentation: Improves segmentation accuracy for tasks that require precise delineation of biological structures in medical and experimental images.
• Quantitative analysis and clinical research: Enhanced segmentation metrics facilitate downstream quantitative measurements, diagnostic interpretation, and studies of biological processes.
• Evaluation on public datasets: Demonstrated metric improvements across seven publicly available biomedical imaging datasets.

Methodology:

The method prunes prediction maps from earlier epochs and integrates the previous-epoch mask with the current-epoch feature map to apply hard attention within convolutional layers, using this feedback mechanism to iteratively refine predictions during training and to perform iterative rectification at test time.