Automatic Detection of Renal Abnormalities by Off-the-shelf CNN Features

Abstract

ABSTRACTThe chronic kidney disease (CKD) and end-stage renal disease are the leading causes of kidney (renal) failure which may be fatal. Thus, it is important to diagnose the kidney diseases at an earlier stage before it cause fatality. This paper proposes an algorithm for automatic kidney abnormalities detection and classification at an early stage. Kidney abnormalities detection involves two stages: feature extraction and detection. The input images used in this work are kidney ultrasound (US) images which are classified into three categories such as normal kidney, kidney with cyst and kidney with stone. The pre-trained convolutional neural networks (CNN) is used to extract features from the kidney US images then, the extracted features from CNN are fed into the support vector machine (SVM) classifier for classifying the kidney abnormalities. The training and testing phase of CNN requires large number of labelled images. Due to unavailability of larger datasets in the kidney US images, off-the-shelf CNN features are used which perform well for smaller datasets. Performance evaluation is done for classification and 91.8% of the accuracy is obtained. Further, the sensitivity and specificity achieved for the SVM classifier in detecting normal kidney images is 93.75% and 92.5%, respectively.