UNet++: A Nested U-Net Architecture for Medical Image Segmentation.

Area of Science:

• Computer Vision
• Medical Imaging
• Deep Learning

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Existing U-Net architectures face challenges in reducing the semantic gap between encoder and decoder feature maps.
• Improved segmentation models are needed for diverse clinical applications.

Purpose of the Study:

• To introduce UNet++, a novel, deeply-supervised encoder-decoder network for enhanced medical image segmentation.
• To address the semantic gap between encoder and decoder feature maps through redesigned nested, dense skip pathways.
• To evaluate the performance of UNet++ against established U-Net architectures across various medical imaging tasks.

Main Methods:

• Developed UNet++, a deeply-supervised encoder-decoder network.
• Implemented nested and dense skip pathways to bridge the semantic gap.
• Conducted comparative evaluations on chest CT nodule, nuclei, liver, and polyp segmentation tasks.

Main Results:

• UNet++ demonstrated superior performance compared to U-Net and wide U-Net.
• Achieved an average Intersection over Union (IoU) gain of 3.9 points over U-Net.
• Attained an average IoU gain of 3.4 points over wide U-Net, showcasing significant improvements.

Conclusions:

• UNet++ offers a more powerful and effective architecture for medical image segmentation.
• Deep supervision combined with redesigned skip pathways facilitates easier optimization and better results.
• The proposed architecture shows broad applicability across multiple challenging medical imaging segmentation tasks.