COVID-19 Detection from Pulmonary CT Images using Neural Networks based on Dropout-Driven Hidden Layers

Enhanced diagnosis with considerably good sensitivity and specificity is highly indispensable for COVID-19 diagnosis using radiological data to combat hazardous viral infection. Accuracy of diagnosis is a very important part that helps in further triaging and disease management. Artificial intelligent techniques using Convolutional Neural Networks and their modified alternatives have been recognized to be the salvation in chaotic situations and emergencies. Despite their immense ability to give quality results, they suffer from overfitting problems which have to be reduced by regularizing the networks. Dropout is one such regularization that modifies the network to achieve improved performance by discarding the unwanted nodes in the network layers. A simple neural network architecture inspired by former renowned architectures with dropout-driven hidden layers, CVDNN is built and experimented with for various dropout probabilities (0.1, 0.25, 0.5 and 0.75). The model was also tested with different numbers of dense layers: CVDNN1 with a single dense layer and CVDNN2 with two dense layers of a fixed dropout probability of 0.5 in it. The models are trained and tested with pulmonary computed tomography images to distinguish COVID-19 abnormality against normal cases. The CVDNN2 model presents better functioning with improved performance measures than CVDNN1 with an accuracy of 92.86 % accuracy, 90.21% sensitivity and a specificity of 95.52% for the dataset used. Dropout probabilities of 0.25 and 0.5 present reliable and better results compared to the other values experimented with. Hence a dropout-driven hidden layer can enhance the neural network's performance by choosing either 0.25 or 0.5 preferably for different applications.