Um método de dois estágios para detecção de pólipos em imagens de colonoscopia usando aprendizado profundo

Abstract

The gastrointestinal tract, the pathway responsible for the entire digestive process, can be affected by various types of diseases, including colorectal cancer, which is the third leading cause of cancer death. Colorectal polyps, benign tumors detectable through images captured by colonoscopes in the colon of the large intestine, are its main precursors. However, many polyps are overlooked during colonoscopy due to technical and cognitive challenges. Studies indicate that improving the detection rate of these lesions can significantly reduce the risk of colorectal cancer. Therefore, computer-assisted detection techniques are being developed to enhance detection quality during regular exams. This thesis presents a two-stage polyp detection method for colonoscopy images. The first stage involves identifying possible polyp areas using a saliency map extraction model supported by the extracted depth maps. Depth maps are images that represent the distance between objects and the camera, and saliency maps are images that highlight the most relevant regions for human visual perception. Initially, the scope area is reduced by extracting the salient objects (S) present in the image so that this segmented area is combined with the green (G) and blue (B) channels of an image in RGB standard, thus forming a new 3-channel image, known as SGB. The second stage of the method consists of detecting polyps in the extracted images resulting from the first stage, combined with the green and blue channels. For this, models based on convolutional neural networks and Transformers were applied, which are deep learning techniques capable of extracting complex visual features and making accurate classifications. Several experiments were carried out using four public colonoscopy datasets. The best results obtained for the task of polyp detection were satisfactory, achieving 91% Average Precision on the CVC-ClinicDB base and 92% Average Precision on the Kvasir-SEG base, both with SGB images, using the Transformers architecture. In addition, the proposed method outperformed others in the literature on some bases.