Development and evaluation of USCnet: an AI-based model for preoperative prediction of infectious and non-infectious urolithiasis.

Abstract

Background: Urolithiasis, a prevalent condition characterized by a high rate of incidence and recurrence, necessitates accurate preoperative diagnostic methods to determine stone composition for effective clinical management. Current diagnostic practices, reliant on postoperative specimen analysis, often fail to facilitate timely and precise therapeutic decisions, leading to suboptimal clinical outcomes. This study introduces an artificial intelligence model developed to predict infectious and non-infectious urolithiasis preoperatively using clinical data and CT imaging.

Methods: Data from December 2014 to November 2021 involving 642 patients undergoing surgical treatment for urolithiasis were used to train and validate the model. The model integrates Visual and Textual Transformation (VTT) and Multimodal-Segmentation Attention Fusion (MSAF) modules to enhance the diagnostic process.

Results: The model demonstrated superior accuracy and reliability in differentiating between infectious and non-infectious urolithiasis compared to traditional diagnostic methods. It achieved a classification accuracy of 79.66%, Area Under Curve of 86.74%, significantly outperforming conventional ResNet architectures and similar models. The inclusion of clinical parameters substantially improved the model's predictive capabilities.

Conclusions: Our model provides an efficient tool for the preoperative identification of urolithiasis type, supporting clinical decisions regarding surgical planning and postoperative care. Its ability to process and analyze complex clinical and imaging data preoperatively positions it as a valuable adjunct in urological practice, particularly in settings with limited access to specialized medical resources.