A Novel Weighted Average Ensemble Deep Learning Architecture to Identify Pneumonia Diseases from Multi-Modal Images

COVID-19 is a novel virus that infects human’s upper and lower respiratory systems. The COVID-19 pandemic in 2020 and 2021 resulted in staggering cases and deaths, with scientists and health professionals still seeking the most effective containment methods. In this circumstance, conducting timely and cost-effective screenings of affected individuals can help combat this condition. This research describes a Deep Convolutional Neural Network-based approach for identifying COVID-19 patients using chest radiographs. The suggested study includes a number of cutting-edge CNN models, including VGG16, ResNet50V2, and DenseNet121, as well as four additional layers for each transfer learning. Next, the models are integrated using an optimal weighted average ensembling technique to forecast COVID-19. We propose an algorithm to discover the optimal set of weights to ensemble the modified pre-trained model. The CNN models were fine-tuned using a transfer learning technique to improve training and validation accuracy. Partial data augmentation procedures are utilized to expand the training dataset in a balanced manner. Our suggested method achieved a classification accuracy of 96.38%, 99.81%, and 100% for the CXR, CT, and Ultrasound images, respectively; all of these are superior to current CNN models and the benchmark algorithm.