Polyp Segmentation in Colonoscopy Images using U-Net-MobileNetV2

Abstract: Colorectal cancer from the appearance of polyps that can be benign or malignant is one of the most fatal diseases in the world. To find these polyps in patients, colonoscopy is performed, which is a very efficient technique in this case. Clinically, detecting and segmenting these polyps in order to determine their presence or not is a very difficult process that requires a lot of time and experience from professionals, depending directly on these factors. Therefore, it becomes increasingly important to have an automatic, effective and reliable method of detecting and segmenting these polyps, making diagnoses faster and more accurate. In order to assist in the development of a method, we proposed the U-Net-MobileNetV2 model, which is the combination of two neural networks, where one acts as an encoder for the other and is responsible for learning image resources. Our experiments generated satisfactory results, demonstrating a good performance and good segmentation. U-Net-MobileNetV2 achieved a Dice Coefficient of 89.71% and an IoU of 81.64% for the Kvasir-SEG dataset, where both are higher than the results obtained by other state-of-the-art models.