| Abstract |
Introduction: The implementation of artificial intelligence in healthcare has revolutionized the way in which diseases are dealt with, improving the life expectancy of patients, thanks to the development of an ecosystem where new techniques, technologies, patients, and doctors are integrated for a correct diagnosis and treatment. The connection between technology and healthcare is constantly growing, evolving, and optimizing. The coronavirus disease (COVID-19) has been a sickness that has caused multiple effects on our society, from the health sector to the economic sector; this disease has impacted the health of millions of people around the world and may continue to be affected today. The use of tools such as computed tomography (CT) images and deep learning algorithms such as convolutional neural networks are useful for detecting, quantifying, and optimizing the diagnosis of different pathologies, and in this case, it can also be valuable for the quantification of the condition of patients with lung lesions caused by COVID-19. In this work, the U-Net convolutional neural network architecture is implemented with the aim of segmenting regions with abnormalities present in CT images associated with COVID-19. Methodology: A database of a group of patients with a positive diagnosis of COVID-19 was used, which had 2581 CT images of the thoracic region with their respective segmentation masks indicating the lung area where the lesion occurs. For the training of the model, a personalized loss function was implemented, which takes into account the Sorensen-Dice similarity coefficient (DICE) and the categorical cross-entropy loss function for the adjustment of the weights of the network. In addition, a function was programmed to variably adjust the learning rate used by the model during training. The maximum number of epochs for training was set to 200. Results: The training strategy implemented allowed obtaining a model with an average performance in the evaluation metrics DICE score and Jaccard index of 0.892 and 0.789, respectively, using images from the test set. This type of model can be incorporated to support the diagnosis and monitoring of the evolution of COVID-19 in the health services of a smart city, being a tool capable of offering a quick result with reliable performance, which can help the work of the doctor who performs the analysis of the computed tomography studies of patients with this disease. © 2024, The Author(s), under exclusive license to Springer Nature Switzerland AG. |
