FANet: A Feedback Attention Network for Improved Biomedical Image Segmentation.

Area of Science:

• Biomedical image analysis
• Machine learning
• Computer vision

Background:

• Large clinical and experimental datasets are advancing biomedical image analysis.
• Image segmentation is critical for quantitative analysis in this field.
• Deep learning models show promise but underutilize information across training epochs.

Purpose of the Study:

• To develop a novel deep learning architecture for biomedical image segmentation.
• To effectively leverage information from previous training epochs to enhance current predictions.
• To improve the accuracy and robustness of image segmentation models.

Main Methods:

• Proposed a Feedback Attention Network (FANet) architecture.
• Integrated previous epoch masks with current epoch feature maps using attention mechanisms.
• Enabled iterative prediction refinement during testing.

Main Results:

• FANet demonstrated substantial improvements in segmentation metrics across seven public biomedical imaging datasets.
• The feedback attention mechanism effectively pruned prediction maps and rectified segmentation errors.
• The model showed significant gains in segmentation performance compared to existing methods.

Conclusions:

• The proposed FANet architecture significantly enhances biomedical image segmentation accuracy.
• Leveraging historical epoch information via feedback attention is a promising strategy for deep learning models.
• FANet offers an effective and robust solution for quantitative biomedical image analysis.