Abstract
To aid in the diagnosis, monitoring, and surveillance of bladder carcinoma, this study developed a vision-based algorithm that creates referenceable 2D bladder maps from standard cystoscopy videos without specialized hardware. By sequentially stitching image frames based on matching surface features (SURF and RANSAC), the system generates a comprehensive map and enables real-time localization during follow-up procedures. Evaluated in 2D models, 3D phantoms, swine models, and retrospective clinical data, the algorithm demonstrated high feature-matching reliability (93-99% of frames) and rapid localization (under 5 seconds). This tool offers a cost-effective method for standardized documentation and longitudinal tracking of bladder lesions.
Introduction: The Challenge of Longitudinal Bladder Surveillance
Bladder cancer is characterized by high recurrence rates, necessitating frequent and lifelong surveillance via cystoscopy. Currently, urologists document findings using subjective 2D hand-drawn diagrams or raw video files. These methods lack spatial context, making it difficult to precisely revisit specific lesion sites during subsequent procedures. While advanced 3D reconstruction and electromagnetic tracking systems exist, their high cost and requirement for specialized hardware limit their accessibility, particularly in low-resource settings. There is a clear need for a hardware-agnostic, software-based solution that leverages existing video feeds to create a spatial record of the bladder.
Methodology: Vision-Based Stitching and Localization
The proposed pipeline consists of two primary components: a stitching algorithm for map generation and a "revisit" algorithm for localization.
Feature Detection and Map Generation
The system utilizes Speeded-Up Robust Feature (SURF) to identify stable landmarks, such as blood vessel bifurcations, within each video frame. To establish spatial relationships, Random Sample Consensus (RANSAC) calculates the affine transformation between sequential frames, accounting for rotation, translation, and zoom.
To mitigate "drift"—the accumulation of small errors over long sequences—the authors implemented a drift correction step. This involves identifying similar features between spatially adjacent but temporally distant frames, allowing the algorithm to "close the loop" and maintain global map accuracy.
The Revisit Logic
Once a reference map is established, the algorithm can localize the cystoscope during a follow-up exam. By comparing new video frames against the existing map, the system identifies matching features to determine the scope's current position and orientation in near real-time.
Evaluation and Results
The algorithm was validated across four scenarios: a 2D benchtop model, a 3D anthropomorphic phantom, a preclinical swine model, and retrospective human clinical video.
Performance Metrics
- Feature Sufficiency: Between 93% and 99% of video frames contained sufficient features (≥6 matches) for successful tracking, confirming that natural bladder vasculature provides an adequate "fingerprint."
- Localization Speed: In the 3D phantom, the system localized the initial frame of a new sweep within 4.5 seconds, with subsequent frames processed in approximately 0.35 seconds.
- Clinical Feasibility: The algorithm successfully processed archived human cystoscopy data, generating a high-resolution composite map (9837×8585 pixels) in 168 seconds.
- Hardware Agnosticism: The system performed reliably across various devices, from low-cost disposable flexible scopes (360p) to high-definition rigid surgical scopes (720p).
Discussion and Future Directions
By shifting the complexity from hardware to software, this algorithm provides a scalable path for advanced surgical navigation. The resulting maps offer a "ground truth" that reduces inter-operator variability and improves the precision of longitudinal monitoring.
However, limitations remain. The bladder is a non-rigid organ; variations in fluid distension can warp the mucosal surface, potentially complicating feature matching. Future work will focus on incorporating elastic registration to account for these deformations and conducting prospective clinical trials to evaluate the tool's impact on active surgical workflows and patient outcomes.
Summary updated