Introduction: In recent years, computer vision and Deep Learning have been used to detect many different diseases and lesions in the body automatically. Some examples are: Detection of tumor types in lungs, breast, head and brain; diabetic retinopathy classification, skin lesion classification. In previous study, the number of stones was too small and only stones of single composition were studied. In our study, we tried to predict the composition of stones based on a large number of stones and stones of mixed components.
Objectives : To evaluate the accuracy on predicting the components of the urinary stones with the photographic image of stones using deep learning.
Methods: From January 2018 to March 2021, total 1,332 stones were obtained through endoscopic surgery. The components of each stone fragment were verified with conventional stone analysis. These 1332 stones were analyzed for components through stone analysis, and as a result of classification by component, a total of 32 classes were classified. The digital photograph images of all urinary stones were trained by deep learning, and there were top 4 classes that produced significant results out of a total of 32 classes. The top 4 classes were included in the deep learning, and the number of stones in each class was more than 100; Class A: Calcium oxalate monohydrate; Class B: Calcium oxalate monohydrate and struvite; Class C: Calcium oxalate monohydrate and calcium oxalate dihydrate; Class D: Uric acid. Images of the stones were captured on a digital camera and the AI was trained with these images. We analyzed the sensitivity and specificity for each class and each stone component.
Results: A total of 965 urinary stones were included. Xception_Ir0.001 was used in the final analysis which demonstrated the highest accuracy. The overall accuracy of predicting the composition of stones was 91%. The sensitivity for each class was as follows: Class A(94.24%), Class B(85.42%), Class C(86.86%), Class D(94.96%). The sensitivity and specificity predicted for each stone component were as follows: Calcium oxalate monohydrate (98.82%, 94.96%), struvite (85.42%, 95.59%), calcium oxalate dihydrate (86.86%, 99.64%) and uric acid (94.96%, 98.82%).
Conclusions: The method of predicting stone components from raw images using deep learning demonstrated a high accuracy. It was able to classify the components of stones with deep learning whether they were single or complex components. This might be an alternative tool for conventional stone analysis and help physicians to predict components of urinary stones to improve further stone treatment.