Biomedical image segmentation algorithm based on dense atrous convolution.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Biomedical images exhibit complex tissue structures with inter-individual variations.
• Current deep learning methods struggle with accurate segmentation of biomedical images, especially those with significant target variations, leading to missegmentation and missed targets.

Purpose of the Study:

• To develop an advanced biomedical image segmentation method that enhances accuracy and robustness.
• To address the limitations of existing deep learning models in segmenting complex and variable biomedical image data.

Main Methods:

• Proposed a U-Net based network incorporating a dense atrous convolution (DAC) module for multi-scale feature extraction.
• Introduced a dense residual pooling module to further enhance multi-scale feature detection.
• Implemented an attention mechanism in the decoding path to suppress background noise and focus on target areas.

Main Results:

• The proposed method demonstrated superior segmentation performance compared to mainstream networks on biomedical images with diverse targets.
• Significantly reduced instances of missed and missegmented regions.
• Achieved higher segmentation accuracy, yielding results closer to ground truth.

Conclusions:

• The novel dense atrous convolution approach effectively improves biomedical image segmentation accuracy and robustness.
• This method offers a promising solution for accurate segmentation of challenging biomedical images with complex and variable structures.