Deep Learning-based Respiratory Anomaly and COVID Diagnosis Using Audio and CT Scan Imagery

Respiratory diseases and diverse lung conditions could progressively become fatal. Early diagnosis is essential in increasing chances of survival. In this research, we conduct respiratory anomaly and COVID-19 diagnosis using both audio and CT scan imagery inputs. Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, which show impressive performance for diverse image and audio classification tasks, are adopted in this research for the development of diagnostic tools to promote early lung condition diagnosis. Specifically, we propose a bidirectional LSTM (Bi-LSTM) model for respiratory anomaly detection from audio signals. Three deep networks, i.e. EfficientNet-B0, EfficientNet-B3, and EfficientNet-B7, are employed for the identification of COVID-19 from CT scan images. A grid search optimisation is also proposed to conduct optimal hyper-parameter identification for all the spatial and temporal deep networks. Evaluated using both ICBHI respiratory audio dataset and the lung CT scan imagery COVID-CT database, the proposed deep networks show impressive performance for the identification of diverse chronic and non-chronic lung diseases (e.g. COPD and Pneumonia) as well as COVID conditions. The empirical results indicate that the best performance has been achieved by the EfficientNet-B7 model for the COVID-CT dataset with sensitivity and specificity scores of 0.945 and 0.952 respectively.