Modeling a Fine-Tuned Deep Convolutional Neural Network for Diagnosis of Kidney Diseases from CT Images

Abstract

Kidney abnormalities are common in the world and due to the scarcity of urologists, there is a need to develop an automated system for the detection of kidney diseases. In this paper, we present a powerful and unique deep transfer learning (TL) architecture based on a pre-trained VGG19 model and a naïve Inception module to efficiently detect major kidney diseases from CT images. We customized the architecture of the VGG19 model by removing the fully connected layers and placing a randomly initialized naïve Inception module and other layers such as batch normalization, dropout and dense layers to avoid vanishing gradient and overfitting problems and to improve performance. We consider two ways of TL: feature extractor and fine-tuning, which is technically implemented and evaluated. Grad-CAM and feature map techniques are used to visualize regions of interest that are meaningful for the diagnosis of kidney disease. 4000 kidney CT images were used to evaluate the performance of the proposed model using different performance metrics. Experimental results show that the proposed model exhibits superior performance after applying fine-tuning compared to other method. The results showed that the proposed model with a fine-tuning strategy produced an excellent accuracy of 99.25% in detecting three different kidney diseases from CT images.

Topics & Concepts