(E-186) Innovative AI-Driven Analysis of Endoscopic Videos for Identifying Kidney Stone Precursors using Active Learning

Kidney stones, known as renal calculi, pose a significant urological challenge with high recurrence rates. Preliminary stages of kidney stone formation involve the proposed theory of kidney stone precursors, including structures such as Randall’s plaque and collecting duct plugs. Understanding the pathogenesis of kidney stone precursors is crucial for effective prevention and treatment. Endoscopic video analysis provides valuable insights into these precursors’ composition and distribution, but the manual process of identifying and segmenting these precursors is time-consuming and prone to error.

The present study evaluated the performance of SEGI (Segmentation Engine Guided Interface), a computer vision AI model designed explicitly for instance segmentation of kidney stone precursors in endoscopic video frames leveraging the Intel® Geti™ platform, an intuitive user interface based on an active learning approach. The platform intelligently selects images from the pool of unlabeled data to be annotated by the user, incrementally increasing the model’s accuracy while limiting annotation effort.

The SEGI model was trained using extracted frames from kidney endoscopy videos, then manually labeling specific regions of interest, including papilla, plug, and plaque. The model exhibited progressive enhancement in generating live predictions with each training cycle. This iterative nature of human-in-the-loop training played a pivotal role in addressing model limitations and analyzing its performance across model versions. The combination of SEGI and the Intel® Geti™ platform presents a compelling, user-friendly solution to automate medical imaging processing, offering valuable insights for early detection and personalized treatment strategies.