AFN-Net: Adaptive Fusion Nucleus Segmentation Network Based on Multi-Level U-Net.

Area of Science:

• Medical Image Analysis
• Computational Biology
• Computer Vision

Background:

• Nucleus segmentation is crucial in medical imaging but challenging due to small targets and complex boundaries.
• Traditional methods struggle with accuracy in complex medical image datasets.
• U-Net architecture offers a promising foundation for image segmentation tasks.

Purpose of the Study:

• To develop a novel and improved nucleus segmentation method for medical images.
• To address limitations of traditional methods in detecting small nuclei and complex boundaries.
• To enhance segmentation accuracy and efficiency using a modified U-Net architecture.

Main Methods:

• Proposed a Weighted Feature Enhancement Unit (WFEU) for adaptive feature map weighting in U-Net.
• Introduced a Double-Stage Channel Optimization Module (DSCOM) to preserve high-resolution information and improve small target segmentation.
• Developed an Adaptive Fusion Loss Module (AFLM) for balanced optimization of segmentation and classification accuracy.

Main Results:

• The novel method achieved high performance on the 2018 Data Science Bowl dataset.
• Achieved an Intersection over Union (IOU) score of 0.8660 and a Dice score of 0.9216.
• Demonstrated significant advantages over state-of-the-art models with a parameter size of only 7.81 M.

Conclusions:

• The proposed method effectively segments small targets and complex boundaries in medical images.
• The enhancements lead to superior segmentation accuracy and region consistency.
• This research offers valuable insights for future medical image segmentation model development with reduced computational cost.