Sharp U-Net: Depthwise convolutional network for biomedical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• U-Net is a popular fully convolutional network for biomedical image segmentation.
• Traditional U-Net architectures suffer from blurred feature maps and inaccurate segmentation due to direct merging of dissimilar features via skip connections.

Purpose of the Study:

• To introduce Sharp U-Net, an improved fully convolutional network architecture for biomedical image segmentation.
• To address the limitations of U-Net, specifically blurred feature maps and over/under-segmentation.

Main Methods:

• Developed an end-to-end depthwise encoder-decoder fully convolutional network named Sharp U-Net.
• Incorporated a novel sharpening filter layer that applies depthwise convolution with a sharpening kernel to encoder feature maps before merging with decoder features.
• This layer fuses semantically similar features and smooths artifacts early in training.

Main Results:

• Sharp U-Net consistently outperformed or matched state-of-the-art baselines across six diverse datasets for both binary and multi-class segmentation.
• The proposed model achieved superior performance without introducing any additional learnable parameters.
• Sharp U-Net demonstrated effectiveness compared to baselines with significantly more learnable parameters.

Conclusions:

• Sharp U-Net offers a simple yet effective solution for improving biomedical image segmentation accuracy.
• The novel sharpening filter mechanism enhances feature fusion and reduces segmentation inaccuracies.
• This architecture provides a parameter-efficient and high-performing alternative to existing U-Net models.