應用機器學習於尿液結晶型態的判讀

林俊廷^1,3、曾一修^1,2

亞東紀念醫院 外科部 創傷科¹與泌尿科²；醫學研究部³

Apply machine learning to the interpretation of urinary crystal morphology

Jun-Ting Lin^1,3, Yi-Shiou Tseng^1,2

Divisions of Traumatology¹ and Urology², Department of Surgery, Far Eastern Memorial Hospital, New Taipei City, Taiwan

Department of Medical Research³, Far Eastern Memorial Hospital, New Taipei City, Taiwan

Purpose:

Urine crystallography is an intuitive method in clinical testing; however, the identification accuracy of urine crystals needs to be determined by a microscopic examination, which is time-consuming and difficult in clinical laboratories. Calcium oxalate (CaOx) is the most common type of urine crystallography, which is divided into monohydrate (COM), octahedron (OTC) and dodecahedron (DOD). We use machine learning to speed up the process of identification for accurate interpretation, even to replace human for identification, and to provide a better diagnostic tool.

Materials and Methods:

The 10 ml morning urine were obtained from nephrolithiasis patients. The 10 µl urine sample was taken and loaded in hemocytometer then use the microscopy to examine crystal morphology, numbers, and size. Image preprocessing and tagging categories, machine iterative learning and inference were done by the software. The final statistics were interpreted and analyzed for the accuracy.

Results:

There were 926 particles divided into four categories including crystals (COM, OCT, DOD) and others (RBC, debris). Kappa value was 0.894 (Kappa between 0.81 and 1 as almost perfect agreement). Accuracy of machine interpretation was 95%. Sensitivity and specificity of machine interpretation were 96% and 94%.

Conclusions:

There was a favorable sensitivity and specificity between machine and human interpretation. This study provides a novel and time saving tool to identify different types of CaOx stones.