Stroke detection and segmentation in CT images using Convolutional Neural Networks and Active Contour Geodesic Method

Abstract

A challenge in computer vision is the aid to medical diagnosis; a standard process in these systems is the segmentation of diseases in medical images, such as computed tomography (CT) scans. Stroke stands out in several countries, being one of the leading causes of death globally. The creation of artificial intelligence systems can aid in diagnosing diseases using CT scans to segment and extract useful information. In this work, we propose a solution for the segmentation of CT images using Convolutional Neural Networks and techniques of Digital Image Processing called Geodesic Active Contour Method, aiming to improve the aid in medical diagnosis. We use Detectron2 as a neural network to perform the primary segmentation of the stroke. Then, a post-processing of the network outputs is performed using the Geodesic Active Contour method. We obtained outstanding results in segmentation with this methodology, such as accuracy reaching above 99%. Our method aims to bring a refined solution for the segmentation of medical exam images using deep learning and computer vision techniques.