Utilizing adaptive deformable convolution and position embedding for colon polyp segmentation with a visual...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Colorectal cancer (CRC) diagnosis relies heavily on expert polyp detection, a complex task.
• Automated systems aid early CRC diagnosis, reducing time and errors.
• Deep learning segmentation models are crucial for automated CRC diagnosis, with Vision Transformers (ViTs) showing promise.

Purpose of the Study:

• To introduce Polyp-Vision Transformer (Polyp-ViT), a novel deep learning model for enhanced polyp segmentation.
• To improve upon existing Vision Transformer architectures by incorporating adaptive mechanisms for feature extraction and positional embedding.
• To evaluate the performance of Polyp-ViT on benchmark datasets for polyp segmentation.

Main Methods:

• Development of Polyp-Vision Transformer (Polyp-ViT), a novel Transformer model.
• Enhancement of the conventional Transformer architecture with adaptive mechanisms for feature extraction and positional embedding.
• Testing and validation of Polyp-ViT on the Kvasir-seg and CVC-Clinic DB datasets.

Main Results:

• Polyp-ViT achieved high segmentation accuracies: 0.9891 ± 0.01 on Kvasir-seg and 0.9875 ± 0.71 on CVC-Clinic DB.
• The model outperformed existing state-of-the-art models in polyp segmentation tasks.
• Demonstrated the effectiveness of adaptive mechanisms and enhanced positional embeddings in Vision Transformers for medical imaging.

Conclusions:

• Polyp-ViT is a highly effective tool for polyp segmentation in colorectal cancer diagnosis.
• The model's performance suggests its potential as a prospective tool for various medical image segmentation tasks.
• Polyp-ViT's generalizability indicates its adaptability to other medical imaging challenges.