Comparative Deep Learning Framework for Kidney Disease Detection Using CT Imaging with Smart Hospital Recommendation

Abstract

Kidney disorders like tumors, cysts, and stones are a major health issue in the world, which in most cases prove life threatening unless timely diagnosis and treatment are administered. A major imaging modality that is used to evaluate these conditions is Computed Tomography (CT), but manual radiologist interpretation is both time consuming and prone to diagnostic variability. The paper provides a proposal of an automated end-to-end system that integrates the principles of deep learning to classify kidney CT images into four: normal, cyst, tumor, and stone. We performed a strict comparative evaluation of four Convolutional Neural Network (CNN) models namely MobileNetV2, ResNet50, DenseNet121 and VGG16 optimized using Adam, RMSprop and Stochastic Gradient Descent (SGD). As it was proved in our experiments, when trained using the Adam optimizer, the ResNet50 model resulted in a maximum classification accuracy of 99% and outperforms all other variants. In order to bring this diagnostics capacity into a practical clinical use, we created a responsive web application with React.js. Using this platform, the user is able to take out a CT scan and once a classification has been taken, the system uses user geolocation information to suggest the top three closest specialized healthcare facilities. This combined solution does not only automate the kidney pathology detection process with great accuracy, but also provides a critical gap that exists between the diagnosis and access to prompt medical attention, a useful early intervention tool.