Overactive Bladder (OAB) is an increasingly prevalent urological condition. Routine Urodynamics Study (UDS) is invasive, inconsistent and simply impractical in every outpatient setting. We propose a novel machine vision augmentative model on cystoscopic images to identify segments of Detrusor Overactivity (DO) based on motion differences of the vascular network across time.
We prospectively collected 158 cystoscopy videos [38 previously UDS confirmed DO, 120 non-OAB] and extracted 30 seconds clip from each video, with an image capture rate of 24 frames/sec, amounting to >100,000 frames. The recordings were first denoised to remove artefacts and account for the cystoscopic movements, and have the resolutions enhanced with a set of Deep Learning techniques. Following which, the vascular network is segmented out and multiple key points are determined along the visualised blood vessels. An average of ~300 key points were identified on each frame captured. Next, mosaic stitching is performed to reconstruct the extracted 2D images into a 3D bladder map to account for the geometric distortions. Finally, the motion differences of the key points across timeframes are tracked as a surrogate for areas of detrusor microcontraction.
The structure-from-motion pipeline demonstrated satisfactory 3D reconstructions of the processed cystoscopy videos. The AI-based semantic image segmentation was trained to differentiate UDS confirmed DO from non-OAB patients with a coefficient score of 0.613. The motion differences could be tracked both quantitatively (number of key points) and qualitatively (extent of displacement of key points over time). Areas of exaggerated motion differences detected would represent increased focal detrusor contractions.
This novel machine vision augmentation model yields promising results in identifying OAB patients based on cystoscopy. Potential identification of detrusor microcontraction points can aid targeted drug delivery to enhance treatment efficacy.
