Presentation Authors: Ali H Aldoukhi, Ann Arbor, MI, Hei Law, Princeton, MI, Kristian M Black*, William W Roberts, Ann Arbor, MI, Jia Deng, Princeton, MI, Khurshid R Ghani, Ann Arbor, MI
Introduction: Deep learning is a subset of machine learning and artificial intelligence in which algorithms are developed to train machines/computers to detect a specific feature or classify objects. There is limited assessment of this method to identify and classify images in urology. We sought to assess the accuracy of deep learning method to automatically detect stone composition from images of kidney stones.
Methods: Human kidney stones of different compositions were obtained from a stone analysis laboratory including pure (>90%) Calcium Oxalate Monohydrate (COM), Uric Acid (UA), Magnesium Ammonium Phosphate Hexahydrate (MAPH), Calcium Hydrogen Phosphate Dihydrate (CHPD) and Cystine stones. Two images were captured for the majority of stones and images were cropped to remove the background. A deep convolutional neural network was trained with the images to predict the stone composition. The outcome was accuracy of the network in detecting stone composition by cross validation method.
Results: A total of 63 stones were used in this study including 21 COM, 17 UA, 7 MAPH, 14 CHPD, and 4 Cystine stones consisting of a total of 128 images. The accuracy of identifying stone composition were as follow: 90% for COM, 94% for UA, 86% for MAPH, 71% for CHPD, and 75% for Cystine stones. Figure 1 depicts samples of the stones before cropping. Figure 2 demonstrates a schematic for the algorithm used in this study.
Conclusions: Deep learning computer vision methods can be used to detect kidney stone composition with high accuracy and has the potential to replace laboratory analysis of stone composition. The accuracy can be improved by increasing the number of stones used for training the network. Future work is needed to see if deep learning can be used for detecting mixed kidney stone composition.