PolypSegNet: A modified encoder-decoder architecture for automated polyp segmentation from colonoscopy images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer is a leading global cause of death.
• Early polyp detection, crucial for survival, is challenged by variations in polyp size, shape, and subtle visual differences.
• Traditional automated segmentation methods using architectures like U-Net have limitations, leading to suboptimal performance.

Purpose of the Study:

• To propose a novel deep neural network, PolypSegNet, for highly accurate automated segmentation of polyps in colonoscopy images.
• To overcome the limitations of conventional architectures for precise polyp detection and diagnosis.

Main Methods:

• Developed PolypSegNet, an encoder-decoder architecture integrating sequential depth dilated inception (DDI) blocks for multi-scale feature aggregation.
• Implemented a deep fusion skip module (DFSM) for enhanced skip connections between encoder and decoder layers.
• Introduced a deep reconstruction module (DRM) for joint optimization of multi-scale decoded feature maps.

Main Results:

• PolypSegNet achieved high performance across four public databases: 91.52% (CVC-ClinicDB), 92.8% (CVC-ColonDB), 88.72% (Kvasir-SEG), and 84.79% (ETIS-Larib).
• The model demonstrated accurate segmentation of polyp regions, even in challenging visual conditions.
• Mean five-fold cross-validation dice scores indicate superior segmentation capabilities.

Conclusions:

• PolypSegNet offers a significant advancement in automated polyp segmentation for colonoscopy images.
• The proposed architecture effectively addresses limitations of traditional methods, improving diagnostic accuracy and speed.
• This technology has the potential to expedite colorectal cancer diagnosis and improve patient outcomes.