Dual encoder-decoder-based deep polyp segmentation network for colonoscopy images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer prevention relies on colonoscopic polyp detection, a process prone to human error and labor intensity.
• Automated polyp detection and segmentation using deep learning, particularly convolutional neural networks, offer a promising solution to enhance diagnostic accuracy.
• Existing methods face challenges including model overfitting, poor boundary definition, and limited ability to capture polyp variations in texture, size, and color.

Purpose of the Study:

• To develop an advanced deep learning model for accurate polyp segmentation in colonoscopy images.
• To address limitations of current methods such as overfitting, generalization issues, and variability in polyp appearance.
• To introduce the Polyp Segmentation Network (PSNet) as a novel dual encoder-decoder architecture for improved polyp detection.

Main Methods:

• Proposed a novel dual encoder-decoder architecture named Polyp Segmentation Network (PSNet).
• Integrated various deep learning modules, including PS encoder, transformer encoder, PS decoder, enhanced dilated transformer decoder, partial decoder, and a merge module.
• Conducted comparative studies on 5 existing polyp datasets and a newly modified polyp dataset.

Main Results:

• PSNet achieved state-of-the-art performance on 5 existing polyp datasets, with mDice of 0.863 and mIoU of 0.797.
• On a modified polyp dataset, PSNet attained superior results with mDice of 0.941 and mIoU of 0.897.
• Demonstrated significant improvements in polyp segmentation accuracy and robustness.

Conclusions:

• PSNet effectively addresses challenges in polyp segmentation, outperforming existing methods.
• The proposed architecture shows high potential for improving automated polyp detection in clinical settings.
• Enhanced segmentation accuracy contributes to more reliable colorectal cancer screening and prevention strategies.