Urine Crystal Classification Using Convolutional Neural Networks

Abstract

This study focuses on classifying different types of urine crystals using Convolutional Neural Networks (CNN). 1100 data samples are collected from medical books and hospitals and divided as training and testing datasets in a 70:30 percentage ratio. To yield an optimized reliability rate in classifying the types of urine crystals, CNN, a deep learning algorithm is used. First, the images underwent preprocessing stage to eliminate noise, to smooth, and to convert it as a binary image. In the segmentation process of the system, some images that contains overlapping urine crystals, indefinite in shape and colorless crystals become major factors and caused these images not to be optimally segmented. Layers of CNN are trained in a way that it can detect patterns from simple to further complex patterns. A convolution examines the entire image in search of information required for greater prediction accuracy. The system’s overall reliability is to be equal in 87.88%. The error rate for classification was often caused by the overlapping of urine crystals in the test image and differences of some urine crystals in terms of its shape and appearance.