Automated Machine Learning Approach for COVID-19 Diagnosis Based on Deep Efficient Features from Chest CT Images

Computer-aided diagnostic systems utilizing artificial intelligence tools are viable resources aiding clinicians in decision-making. However, with a surfeit of Machine Learning (ML) algorithms, selecting the most promising model with the right set of hyperpa-rameters is often quite challenging. This challenge might be restricting many crucial developments in COVID-19 research. This paper presents an experimental demonstration of the applicability of Automated Machine Learning (autoML) in the screening of COVID-19 infection using deep features extracted from routine chest Computerized Tomography (CT) scan images. For this purpose, we have proposed an EfficientTPOTNet framework, which is a combination of EfficientNet B7 Convolutional Neural Network as a feature extractor paired with a robust ML pipeline generated using Tree-Based Pipeline Optimization Tool (TPOT), an autoML library. The developed framework is validated and compared with other state-of-the-art feature extraction methods and manually created machine learning pipelines using 2482 (1252 COVID and 1230 Non-COVID) chest CT scan images from the SARS-CoV-2 CT dataset. Moreover, we have used 746 (349 COVID and 397 (Non-COVID) independent chest CT scan images from the COVID-CT dataset to study the generalization ability of the ML pipeline on the unseen statistical distribution of data. The EfficientTPOTNet framework achieved a 10×10 fold cross-validation accuracy of 98.45% and an F 1 -score of 98.64% on the SARS-CoV-2 CT dataset.