Identificação de pólipos através de técnicas de segmentação de imagens

Abstract

Cancer is one of the main causes of death in the world, and one of the most diagnosed and lethal is colorectal cancer. Most cases of colorectal cancer originate from polyps, which are disordered cellular growths of the intestinal mucosa. Therefore, one of the ways to prevent colorectal cancer is the identification and removal of polyps through colonoscopy exams. However, a significant portion of polyps may be unnoticed by specialist doctors when analyzing images of this examination, reaching 27% of polyps less than 5 mm. In this way, computer vision algorithms can help these professionals in the identification and delimitation of polyps, especially algorithms that are based on deep learning. Deep learning algorithms learn from a large set of data to recognize valid patterns, in this case, patterns that characterize a polyp in a particular area of an image. Two of the main computer vision tasks that do use of deep learning techniques are object detection and segmentation. The first task delimits the objects of interest using a box, and the second one surrounds precisely each of the objects. In this sense, the objective of this work was to identify labeled public colonoscopy image bases, and to train detection and segmentation algorithms. Two algorithms were used, being YOLOv8 as a detector and segmenter, and SAM as a segmenter. The two algorithms had their performance evaluated separately and together, so that YOLOv8 outputs were used to identify the area to be segmented by SAM. Among the different variants of YOLOv8, the one that presented the best results for the detection task was nano, reaching 0.969 of mAP50 and 0.796 of mAP50-95. In the case of the segmentation task, the best results were reached by YOLOv8 alone (without using SAM) in its large version, reached up to 0.895 for the dice coefficient, and 0.850 of IoU. Although no improvements were observed in the segmentation task combining YOLOv8 and SAM, good results were achieved when the original annotation of the images were used as input, reaching 0.942 of dice and 0.896 of IoU, thus demonstrating the algorithm’s capacity to segment polyps as long as their precise location is provided.