Polyp Segmentation in Colonoscopy Images

Colorectal cancer is a prevalent form of cancer, often detectable through polyps in the gastrointestinal tract. Unfortunately, these polyps typically do not display noticeable symptoms, making early detection challenging. While procedures like colonoscopy and endoscopy can identify polyps, they can miss some, leading to the need for a more automated approach. One innovative solution is capsule endoscopy, which records detailed images of the gastrointestinal tract over an extended period. However, the massive volume of data generated necessitates automation for efficient analysis. Artificial intelligence, particularly convolutional neural networks (CNNs) like TransUNet, can be crucial in quickly and accurately identifying suspicious areas in capsule endoscopy images. This study focuses on automating polyp detection using TransUNet and aims to enhance the early detection of colorectal cancer. The research utilizes the Kvasir-SEG database, containing polyp images and annotated segmentation masks. Various CNN architectures, like UNet, ResUNet, and ResUNet++, are employed, with metrics like Dice Loss and Tversky Loss used for performance evaluation through techniques like cross-validation. Results demonstrate that the TransUNet approach, leveraging transformers in its encoding layers, achieved 66% Dice Score, outperforming other architectures like UNet and ResUNet in this metric, however it did not surpass the ResUNet++ network. In conclusion, the TransUNet model shows potential for automating polyp detection in gastrointestinal images, offering a valuable tool in the fight against colorectal cancer. Integrating advanced technology into medicine promises more accurate and efficient gastrointestinal care.