The aim of this research is to study and develop efficient yet explainable methods to identify COVID-19 via chest X-ray and thoracic computed tomography (CT). While manual reading CT and X-ray images takes 15 minutes and involves a highly skilled medical doctor/consultant which are now in high demand, the use of machine learning can take few seconds on a computer and be automated which provides opportunity for high throughput and remote way of operation.
In this research proposal we suggest the use of a new classification method that is explainable by design and able to continue to learn and adapt for each new data sample. The latter is immensely important for the case of COVID-19 (and other disease), because new cases are being accumulated every minute and traditional approaches require either iterative re-training or ignore the new data. We aim to provide an understandable and interpretable recommendation framework that can be important decision support tool to specialists in the diagnosis and treatment decision-making process.
To train our model we plan to use open source data, as well as, provide our own dataset. The data that we aim to provide is being collected from different hospitals of Sao Paulo,Brazil. We expect to acquire data for 560 patients divided between 280 patients with COVID-19 and 280 non-COVID patients for control. These data will be constituted of thoracic CT images (at least 10 different images per patient) and qualitative data such as age, profession, sex, habits, previous diseases, and others.
The main idea is to provide i) an explainable COVID-19 identification tool via thoracic CT images; ii) a tool to estimate the degree of pulmonary involvement by COVID-19; iii) and a tool predict the risk of dynamic evolution of the pulmonary involvement via data fusion of qualitative features and thoracic CT. In the third case, the explainable deep learning method will be trained having as input the qualitative data provided by a patient and thoracic CT time series (for patients that have done more than one CT-scan in different days) in order to predict the degree of pulmonary involvement for a given n number of days.