NA-segformer: A multi-level transformer model based on neighborhood attention for colonoscopic polyp segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Colon cancer deaths are increasing globally, necessitating improved early detection methods.
• Manual diagnosis of colon polyps is time-consuming and prone to errors.
• Automated computer-aided diagnosis (CAD) using deep learning shows potential for accurate polyp segmentation.

Purpose of the Study:

• To introduce NA-SegFormer, a novel multi-level encoder-decoder architecture for enhanced colonoscopic polyp segmentation.
• To address challenges in segmenting colon polyps, including variations in size and data imbalance.
• To improve the efficiency and accuracy of automated polyp detection in colonoscopies.

Main Methods:

• Developed a Transformer-based segmentation model (NA-SegFormer) with a novel neighbor attention mechanism for patch merging.
• Implemented a unified focal loss function to handle category imbalance in colon polyp datasets.
• Evaluated the model on diverse datasets: Kvasir-SEG, Kvasir-Instrument, and KvasirCapsule-SEG.

Main Results:

• Achieved high performance with Dice scores of 94.30% (Kvasir-SEG), 94.59% (Kvasir-Instrument), and 82.73% (KvasirCapsule-SEG).
• Obtained excellent accuracy rates: 98.26% (Kvasir-SEG), 99.02% (Kvasir-Instrument), and 81.84% (KvasirCapsule-SEG).
• Demonstrated a fast inference speed of 125.01 Frames Per Second (FPS), outperforming existing models.

Conclusions:

• NA-SegFormer significantly improves colon polyp segmentation accuracy and speed compared to state-of-the-art methods.
• The model offers a favorable trade-off between inference speed and accuracy, vital for real-time applications.
• This advancement holds significant promise for real-time colonoscopic polyp segmentation and early colon cancer detection.