Abstract
Ureteral stone is a disease in which crystals form in the urine and form stones in the ureter. In recent years, extracorporeal shock wave lithotripsy (ESWL) and transurethral lithotripsy (TUL) have been established as the principal treatment methods for ureteral stone. ESWL is less physically burden, but its success rate is lower than that of TUL. Physicians select an appropriate treatment based on clinical findings, CT images and so on. However, the success rate of ESWL is still about 70%. The purpose of this study is to reduce the number of patients who suffer the physical and financial burden of double treatment due to ESWL failure through the clinically interpretable and highly accurate predictive model of ESWL outcomes. We collected X-ray and CT images and clinical findings from 162 urolithiasis patients. First, the proposed method extracts image features using deep learning and hand-crafted shape and texture features. Next, it constructs a model to predict ESWL outcomes using machine learning. The proposed model performed with an accuracy of 0.887, a specificity of 0.620, an AUC of 0.890 by using a logistic regression model. We then analyzed the influence of each feature on prediction by using SHAP values. Also, features are selected by using SHAP values to improve the performance. We achieved an accuracy of 0.913, a specificity of 0,857, and an AUC of 0.956 by using random forest where seven features were selected. The experimental results indicated that we were able to reduce the number of features and develop a more interpretable model.
