Lung diseases and COVID diagnosis from computed tomography images using explainable deep learning

Abstract

The use of deep learning for medical decision support over Computed Tomography (CT) and X-Ray images in the current COVID pandemic is a hot topic in medical imaging processing. Research has shown that CT images display signs of COVID infection, which are more prevalent in severe cases. Some methods have already been proposed, mainly using transfer learning techniques over already proven methods, for classification and segmentation of COVID related CT images. Manual annotations are considered to be a gold-standard when developing and evaluating current automated methods, and those are still lacking specially for segmentation. This lack of gold standard data paired with the possibility of overfitting to specific characteristics of the scarce public data currently available puts in check the reliability of those methods. Special attention has to be drawn to lung diseases that present similar findings in CT, and could lead to confusion when evaluated by a method specialized in COVID and NON-COVID cases. In this project, we propose to develop an automated pathology diagnosis method, able to be explainable through two different tasks, classification and accurate segmentation of lung and lung findings in chest CT images, while taking into account the possible presence of other diseases. We intend to build a large cohort of public and private data for evaluation of our method, including more than 3000 classification volumes presenting diverse lung diseases and 124 (possibly growing) labeled segmentation CT volumes, in a partnership with the Radvid19 São Paulo Government initiative and São Paulo hospitals, where we already have promising preliminary results. We also will have access to data from our partnership with the BRAINN (2013/07559-3) and COVID (2020/07200-9) FAPESP CEPID projects, where CT volumes from COVID patients are being acquired from Campinas' Hospital das Clínicas. This data is going through all the necessary ethical approval steps. We plan to use novel segmentation and classification architectures based on recent advances in semantic segmentation, and experimentation with a paradigm shift in medical imaging processing, namely the use of transformers. We will compare our methodology with state-of-the-art techniques in public datasets and subject it to qualitative evaluation from physicians. Plans for a BEPE are part of this PhD project. (AU)