COVID-19 diagnosis with hyperparameter optimization using mixture density network(MDN) and EMADE

ABSTRACT

COVID-19 diagnosis has become a crucial task in today's world due to the rapid spread of the infectious Corona Virus disease caused by the SARS-CoV-2 virus. Analysis of COVID using CT scan images is shown to give better results but it requires expert radiologists and it consumes time. Hence there is a need for a diagnosis system to classify whether it's COVID positive or not for quick and early diagnosis. Deep Learning models are effective in handling classification problems but some models might lead to vanishing gradient problem. A Mixture Density Network (i.e.) Bidirectional Long Short-Term Memory((Bi-LSTM) with Mixture Network is used as the classifier to handle the vanishing gradient problem and to classify based on the probability distribution. Parameter tuning plays a major role in improving the overall efficiency of the classifier. An Enhanced Memetic Adaptive Differential Evolution (EMADE) algorithm is proposed for tuning the parameters of the classifier. Enhanced MADE is a memetic algorithm with proposed Elite chaotic local search (ECLS) which helps in addressing the issue of getting stuck at a local optimal solution and premature convergence. The use of Elitism in the chaotic local search directs the algorithm toward the optimal solution and increases the exploitation ability. Due to high false negatives in RT-PCR, CT scan images have been taken as the input. The dataset is labeled and it consists of 1252 CT scans that are positive for COVID-19, and 1230 CT scans that are negative for COVID-19. The dataset collected from patients in Sao Paulo, Brazil that is available on Kaggle is used [21]. A sample of the dataset is taken for experimentation and an accuracy of 75.83% is achieved. The precision is 80.32% indicating that there are fewer False positive than the existing methods. © 2022 IEEE.