Identification, Segmentation and Localization of GI Polyps in Colonoscopy Using Deep Learning Techniques

Abstract

Computers have improved the old colonoscopy procedures as they are helpful in the early detection of the problem and help in pin-pointing the issue by highlighting the main areas of the interest. With the improvement in computer technology, various computer vision methods have been developed which can be applied on polyp datasets to automatically detect the polyps by outlining its boundaries. But, it is still an on-going challenge to find the best method for the early detection of polyps. Finding the best method for polyps’ detection not only ensures that the results generated by this method are reproducible but also helpful in completing the tasks of Automated identification and segmentation of polyps with greater accuracy. Moreover, it is important to consider that the ideal method should be consistent with industrial standards and should has results that are comparable to other used methods. The purpose of this research is to identify a benchmark method, by analysing several different methods, by publicly available dataset i.e. Kvasir-SEG which contains hundreds of images obtained from colonoscopy that can be used for detection, segmentation and pin-pointing of polyps. By doing this the speed and accuracy of the benchmark method is also evaluated. Most methods which are already proposed in literature perform well but they do not provide accurate results thus these methods cannot be used as a benchmark method. For identification and localization of polyps, ColonSegNet, method has been used in this study which provided satisfactory results with a mean AP and mean IoU of 0.8000 and 0.8100 respectively. This method has the highest speed of 180 frame per second (fps) which is more than enough to prove that the method is useful and can be used as a benchmark method. Not only this, the proposed method has also performed well in segmentation task by achieving 182.38 fps average speed with a comparable 0.8206 dice coefficient. The detail comparison carried out in this study by using different and advanced methods highlights that it is critically important to benchmark deep learning techniques to automate polyps xi identification and segmentation process. The benchmarking can help us to transform clinical practices, which are already in use, to be more accurate and efficient. Thus, the chances of missing polyps’ detection during clinical examination may reduce and a significant improvement in the outcome of clinical examination can be achieved.