Transformer-Dilated ResUNet: An Efficient and Accurate Framework for Polyp Segmentation

Colonoscopy is considered the gold standard test for identifying colorectal cancer (CRC). However, it does have some shortcomings due to the variety of polyps and observer variability. These challenges associated with polyp identification and segmentation have been addressed by the development of deep learning-based systems. Although these systems show promise in terms of accuracy and efficiency, they often struggle with real-time data handling and inter-patient data from various medical facilities. To improve the efficiency of automated polyp identification and segmentation, it is essential to advance real- time capabilities and generalizability. This has the potential to enhance patient outcomes and reduce morbidity and mortality associated with CRC. Several different approaches utilizing computer vision and deep learning have been suggested for polyp segmentation. Tra- ditional methods frequently employ image processing techniques such as thresholding, edge detection, and region growth to identify and segment polyp regions based on color, texture, and shape properties. However, these approaches may have limitations in addressing differences in polyp appearance and complex background structures. In this paper, we propose TransDRN, a new deep-learning architecture for autonomous polyp segmentation. The archi- tecture is based on residual blocks, dilated convolutions, the transformer self-attention mechanism, and dilated convolutions. We demonstrate that TransDRN achieves highly promising re- sults while maintaining real-time processing performance, using benchmark datasets for polyp segmentation accessible to the general population. Due to its superior performance metrics, TransDRN has established itself as a reliable standard for creating real-time polyp detection systems that can aid in early identification, treatment, and prevention of colorectal cancer.