Development of an Improved Mayfly Algorithm Based Convolutional Neural Network for Pulmonary Diseases Recognition System

Abstract

Pulmonary diseases impact the respiratory system. Convolutional Neural Network (CNN) is used for detection and recognition of pulmonary diseases; however, it suffers from hyperparameter selection and overfitting problems. Existing optimization techniques such as the Mayfly Algorithm (MA) also require initial parameter tuning and exhibit slow convergence behaviour. This research developed a Roulette Chaotic Mayfly Algorithm (RCMA) based on CNN (RCMA-CNN) for pulmonary diseases recognition. X-ray images including normal and pulmonary diseases cases were obtained from Kaggle and pre-processed for the desired image quality. The RCMA was formulated using Roulette wheel selection to model attraction deterministically and Chaotic Sinusoidal Map Function to balance exploration and exploitation in the MA. RCMA was applied to optimize CNN hyperparameters including number of layers and batch size at the convolutional layer. This was implemented in MATLAB (R2020a) and compared with MA-CNN and CNN in terms of false positive rate, sensitivity, specificity, accuracy and recognition time. At optimal threshold of 0.75, RCMA-CNN gave false positive rate of 1.43%, sensitivity of 98.06%, specificity of 98.57%, and accuracy of 98.32%. RCMA-CNN recorded a recognition time of 76.81 seconds, which was better than that of MA-CNN and CNN. The RCMA CNN model significantly outperformed both MA-CNN and standard CNN.