COVID-19 Detection and Diagnosis Model on CT Scans Based on AI Techniques

The end of 2019 could be mounted in a rudimentary framing of a new medical problem, which globally introduces into the discussion a fulminant outbreak of coronavirus, consequently spreading COVID-19 that conducted long-lived and persistent repercussions. Hence, the theme proposed to be solved arises from the field of medical imaging, where a pulmonary CT-based standardized reporting system could be addressed as a solution. The core of it focuses on certain impediments such as the overworking of doctors, aiming essentially to solve a classification problem using deep learning techniques, namely, if a patient suffers from COVID-19, viral pneumonia, or is healthy from a pulmonary point of view. The methodology's approach was a meticulous one, denoting an empirical character in which the initial stage, given using data processing, performs an extraction of the lung cavity from the CT scans, which is a less explored approach, followed by data augmentation. The next step is comprehended by developing a CNN in two scenarios, one in which there is a binary classification (COVID and non-COVID patients), and the other one is represented by a three-class classification. Moreover, viral pneumonia is addressed. To obtain an efficient version, architectural changes were gradually made, involving four databases during this process. Furthermore, given the availability of pre-trained models, the transfer learning technique was employed by incorporating the linear classifier from our own convolutional network into an existing model, with the result being much more promising. The experimentation encompassed several models including MobileNetV1, ResNet50, DenseNet201, VGG16, and VGG19. Through a more in-depth analysis, using the CAM technique, MobilneNetV1 differentiated itself via the detection accuracy of possible pulmonary anomalies. Interestingly, this model stood out as not being among the most used in the literature. As a result, the following values of evaluation metrics were reached: loss (0.0751), accuracy (0.9744), precision (0.9758), recall (0.9742), AUC (0.9902), and F1 score (0.9750), from 1161 samples allocated for each of the three individual classes.