#0638

An artificial intelligence system for renal stone imaging diagnosis and prediction of postoperative stone free status

Y. Liu^1,2,3, P. Chiu¹, Y. Liu¹, X. Wang¹, Y. Xie¹, W. Wu¹, Z. Huang¹, K. Huang¹

¹National Kaohsiung University of Science and Technology, Department of Electrical Engineering, Kaohsiung City, Taiwan
²Kaohsiung Chang Gung Memorial Hospital and Chang Gung University College of Medicine, Department of Urology, Kaohsiung City, Taiwan
³Kaohsiung Municipal Hospital - Under the management of Chang Gung Medical Foundation, Department of Urology, Kaohsiung City, Taiwan

Introduction:

Kidney-ureter-bladder (KUB) X-ray image is the common used method for renal stone detection in the emergency room for its low cost and radiation dose. However, accurate interpretation of the KUB images requires experience. This study proposes an artificial intelligence-assisted diagnostic and predictive system that integrates image preprocessing, deep learning, and machine learning to help inexperienced clinicians diagnose renal stones and predict the stone free status after percutaneous nephrolithotomy (PCNL).

Material and methods:

This study collects KUB images and various medical parameters from patients with suspicious renal stones at Kaohsiung Chang Gung Memorial Hospital. Based on this dataset, three subsystems are established (Figure 1). The first subsystem focuses on classifying whether the KUB images contain renal stones. We first calculate the centroid of the image's brightness and use it to crop the image. This step helps to minimize the impact of irrelevant tissues, ensuring that the model focuses on regions where renal stones are likely to appear. After cropping, a Swin Transformer model is used to classify the image as either containing renal stones or not. The second subsystem uses SegViT (Semantic Segmentation with Plain Vision Transformers), for image segmentation of renal stones. First, the model was applied to segment the spine and pelvis. The results were then used to create masks, which were subsequently applied to crop the Region of Interest (ROI). Next, use model to segment the stones in the image. In the third subsystem, we used XGBoost for stone free status prediction after PCNL, utilizing tabular data to predict whether residual stones would remain after surgery.

Results:

This study combines predictions of whether an image contains renal stones (90.6% accuracy), performs renal stone segmentation (87.4% mean Intersection over Union, mIoU), and provides prediction of stone free status after PCNL (91.15% accuracy). We then designed a user interface to display KUB images, highlight stone segmentation, and present the surgical outcomes.