Development of a Computer Vision System for Stone Segmentation During CVAC 2.0 In-Scope Aspiration Ureteroscopy

Technical Summary: StoneClear for CVAC 2.0 Ureteroscopy

Problem Statement
The primary challenge in contemporary endourology is achieving absolute stone-free status during flexible ureteroscopy (fURS). Despite advances in laser lithotripsy and suction-assisted systems like the CVAC® 2.0 Aspiration System, residual stone fragments (RSFs) frequently remain, serving as a nidus for recurrence. A critical limitation is the subjectivity of intraoperative visual inspection. The operative field is often obscured by the "snowstorm" effect of stone dust, blood, and irrigation fluid. Furthermore, the unique environment of suction-assisted ureteroscopy introduces high-velocity particle movement, transient visibility of fragments, and partial occlusion by the suction catheter. Surgeons currently lack an objective, real-time tool to quantify residual stone burden or confirm visual clearance, leading to a discrepancy between intraoperative "visual clearance" and postoperative radiographic stone-free rates.

Methodology
The authors developed StoneClear, an artificial intelligence computer vision system designed to detect and delineate RSFs during Steerable Ureteroscopic Renal Evacuation (SURE) using the CVAC® 2.0 device.

• Data Collection: Intraoperative videos were prospectively collected from 10 patients undergoing fURS with the CVAC® 2.0 device. From over 1.6 million total video frames, a representative sample of 2,416 high-utility frames was selected for dense pixel-level annotation.
• Annotation: Board-certified urologists manually delineated binary stone masks using the MELAGE platform. The dataset included 1,439 stone-containing frames and 977 stone-free frames. Crucially, the stone-free set included challenging background elements such as urothelium, blood, bubbles, and Randall's plaques (fixed mucosal calcifications), which were explicitly annotated as non-stone to train the model to distinguish them from free-floating fragments.
• Model Architecture: StoneClear operates in a two-stage sequential process:
  1. Gatekeeper (Frame-Level Detection): Determines if a stone is present in the frame.
  2. Outliner (Pixel-Level Segmentation): If a stone is detected, it delineates the precise pixel-level boundaries.
     The system utilizes a transformer-based deep learning architecture (PraNet-V2) to extract visual features. Unlike traditional convolutional models that examine local patches, the transformer analyzes relationships across the entire image, aiding in the distinction of true stones from reflections and shadows.
  3. Gating Mechanism: A novel gating mechanism was integrated to suppress false-positive predictions in visually challenging, stone-free frames.
• Evaluation: Model performance was assessed using patient-level 5-fold cross-validation (8 procedures for training, 2 for testing per fold). Metrics included Frame Classification Accuracy, False-Positive Rate (FPR), False-Negative Rate (FNR), and the Dice Similarity Coefficient (DSC).

Key Results

• Detection Accuracy: StoneClear correctly identified the presence or absence of stones in 92.8% of video frames.
• Segmentation Precision: For frames containing visible stones, the model achieved a mean Dice Similarity Coefficient (DSC) of 0.86, indicating excellent spatial overlap between AI predictions and expert surgeon outlines.
• False-Positive Reduction: On stone-free frames, StoneClear produced a false-positive rate of 13.6%. This represents a significant improvement over standard models; without the gating mechanism, the false-positive rate rose to 78.2%, and a U-Net based architecture yielded a rate of 24.75%. The proposed gating mechanism reduced false-positive segmentations by more than fivefold compared to standard segmentation architectures.
• Clinical Context: Preoperative mean stone volume was 10,925.36 mm³, while postoperative mean residual volume was 170.20 mm³.

Key Contributions

1. First CV System for Suction-Assisted Ureteroscopy: This work presents the first computer vision framework specifically developed for the highly dynamic, suction-assisted environment of the CVAC® 2.0 system, addressing a gap where previous intraoperative AI work was limited to non-aspirating fURS.
2. Novel Gating Mechanism: The introduction of a classification-driven gating mechanism effectively mitigates the high false-positive rates typically associated with optical artifacts (blood, bubbles, reflections) in endoscopic video.
3. Robust Proof-of-Concept: The study demonstrates that transformer-based models can achieve high accuracy in distinguishing between treatable residual fragments and fixed anatomical features (like Randall's plaques) in real-time surgical settings.

Significance and Claims
The paper positions StoneClear not as a replacement for postoperative radiologic imaging, but as a robust proof-of-concept for real-time AI-assisted intraoperative visual assessment. The authors claim that by providing objective visual tracking and accurate boundary localization, the framework establishes a critical foundation for estimating intraoperative residual stone volume. This technology aims to assist surgeons in verifying endoscopic clearance before the procedure concludes, potentially reducing the reliance on subjective visual estimation which the study notes often overestimates clearance. The authors draw parallels to successful computer vision applications in gastroenterology (polyp detection), suggesting that AI overlay systems do not require 100% precision to provide meaningful clinical assistance, provided they significantly reduce surgeon oversight.

Limitations
The authors acknowledge the modest cohort size (10 patients) and the single-center nature of the study. Additionally, the evaluation was limited to strategically selected frames rather than the full length of every case due to the intensive labor required for dense, frame-by-frame annotation. Validation on full-length, unselected case recordings is identified as a necessary direction for future work.