Super U-Net: a modularized generalizable architecture.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Classical U-Net architectures face limitations in capturing complex image features.

Purpose of the Study:

• To develop and validate Super U-Net, a novel convolutional neural network (CNN), for enhanced medical image segmentation.
• To evaluate Super U-Net's performance against traditional U-Net and other state-of-the-art models across diverse medical imaging datasets.

Main Methods:

• Super U-Net was developed by integrating dynamic receptive field and fusion upsampling modules into the U-Net architecture.
• The model was trained and validated on fundus, endoscopic, and dermoscopic images for segmenting retinal vessels, GI polyps, and skin lesions.
• Performance was assessed using K-fold cross-validation and metrics including Dice Similarity Coefficient (DSC), accuracy, PPV, and sensitivity.

Main Results:

• Super U-Net achieved superior average DSCs: 0.808 for retinal vessels, 0.752 for pediatric retinal vessels, 0.804 for GI polyps, and 0.877 for skin lesions.
• The novel CNN consistently outperformed U-Net, seven U-Net variants, and two non-U-Net architectures (p < 0.05).

Conclusions:

• The integration of dynamic receptive fields and fusion upsampling modules significantly enhances medical image segmentation performance.
• Super U-Net represents a promising advancement in automated medical image analysis, offering improved accuracy and robustness.