Disease detection by developing an adaptive transfer learning model

Millions of people are affected and die from pneumonia every year around the world. Therefore, early diagnosis of pneumonia is vital for the treatment process. In this study, a web-based diagnosis based on deep learning for the diagnosis of the disease system has been developed. For this purpose, MobileNetV2, ResNet50V2, ResNet152V2, VGG16 and VGG19 deep learning models were used. The fine-tuning strategy was applied by customizing the last layers of deep network models based on the target data. In order for the models to make better predictions, the number of images was increased by using various data augmentation methods. The parameters of accuracy, precision, F1 and Auc-Roc score were used to evaluate the model performance. In experimental studies, the accuracy percentages of MobileNetV2, ResNet50V2, ResNet152V2, VGG16 and VGG19 models were 84.4%, 88.6%, 88.5%, 90%, 87.8%. The accuracy values of the customized models are 89.2%, 92.3%, 87.8%, 91.5%, 89.4%, respectively. When comparing the original models with customized models, the customized ResNet50V2 and customized VGG16 models, exhibited a higher performance with accuracy of %92.3 and %91.5 among the models.