Comparison of deep learning methods in medical image classification

Nowadays, current medical imaging techniques provide means of diagnosing disorders like the recent Covid-19 and Pneumonia due to technological advancements in medicine. However, the lack of sufficient medical experts, particularly amidst the breakout of the epidemic, poses severe challenges in early diagnoses and treatments, resulting in complications and unexpected fatalities. In this thesis study, a CNN model, VGG16 + XGBoost, and VGG16 + SVM, were used for three-class image classification on a generated dataset named Dataset-A with 6,432 chest X-Ray (CXR) images (containing Normal, Covid-19, and Pneumonia classes). Then, pre-trained ResNet50, Xception, and DenseNet201 models were employed for binary classification on Dataset-B with 7,000 images (consisting of Normal and Covid-19). The suggested CNN model achieved a test accuracy of 98.91 %. Then the hybrid models (VGG16 + XGBoost and VGG16 + SVM) gained accuracies of 98.44 % and 95.60 %, respectively. The fine-tuned ResNet50, Xception, and DenseNet201 models achieved accuracies of 98.90 %, 99.14 %, and 99.00 %, respectively. Finally, the models were further evaluated and tested, yielding impressive results. These outcomes demonstrate that the models can aid radiologists with robust tools for early disease diagnoses and treatment.