Machine learning aided kidney stone classification with electromagnetic properties

Electromagnetic waves with microwave frequencies have been widely used in medical field for diagnostic and therapeutic purposes. The reasons why microwaves have been preferred are mainly its non-ionizing nature and its ability to penetrate matter. Physiological effect on biological tissues also enables to use for treatment of diseases. Moreover, the difference in electromagnetic properties of distinct biological material enables diagnostic usage. In this study, the determination of kidney stone types was aimed by making use of the dielectric discrepancy between stone types. It is known from statical information that the recurrence rate of kidney stone disease is very high. For this reason, it is important to learn why the body forms kidney stones in order to determine necessary precautions to prevent recurrence. The tools utilized for determination of the reasons for stone formation are blood test, urine test and tools used for urinary stone analysis. To this end, this study focused on development of an innovative method for urinary stone analysis. When dielectric property measurements of three common types of kidney stone were carried out, it is realized that it is possible to distinguish kidney stones with machine learning algorithms and the help of these features. Dielectric constant and dielectric loss values of each stone were measured over the frequency range of 500 MHz – 6 GHz with 100 MHz intervals. Then, Cole-Cole parameters of each stone were calculated with generalized Newton-Raphson method from measured dielectric constant and dielectric loss values instead of directly use of them. Thus, the input size of machine learning algorithm was reduced with successful Cole-Cole fitting. After dielectric property measurement and Cole-Cole fitting, classification of the stones were completed with an artificial neural network and k-nearest neighborhood algorithm. Obtained classification of two algorithm were evaluated with some performance measures which are accuracy, sensitivity, specificity, precision, recall and F1 score. It can obviously be seen from these measures that both of two classifier are quite successful in determination of kidney stone types.