Multi-level dilated residual network for biomedical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• U-Net is a popular deep learning architecture for biomedical image segmentation.
• Classical U-Net has limitations in learning capability and information loss during feature concatenation.

Purpose of the Study:

• To propose a novel multi-level dilated residual neural network as an extension of U-Net for improved biomedical image segmentation.
• To enhance learning capability and reduce the semantic gap in segmentation tasks.

Main Methods:

• Replaced classical U-Net convolutional blocks with multi-level dilated residual blocks.
• Incorporated non-linear multi-level residual blocks into skip connections.
• Evaluated on five diverse biomedical datasets (electron microscopy, MRI, histopathology, dermoscopy).

Main Results:

• Achieved relative improvements in dice coefficient: 2% (MRI), 3% (dermoscopy), 6% (histopathology), 8% (cell nuclei microscopy), and 14% (electron microscopy).
• Demonstrated robustness against outliers and better boundary continuity compared to U-Net and MultiResUNet.
• Consistently outperformed the classical U-Net across all tested modalities.

Conclusions:

• The proposed multi-level dilated residual U-Net significantly enhances biomedical image segmentation performance.
• The novel architecture effectively addresses U-Net's limitations, offering superior accuracy and boundary preservation.
• This approach shows promise for various medical imaging segmentation applications.