Reducing Fiber-Induced Honeycomb Artifacts and Low-Light Noise in Nasal High-Speed Video Laryngoscopy: A Fast,...

Recordings made with high-speed video laryngoscopy (HSV) in combination with flexible endoscopy are strongly degraded by a combination of fiber-bundle-induced honeycomb artifacts and severe low-light noise, caused by limited illumination through the thin endoscope and low exposure times. More sophisticated denoising approaches, for honeycomb and low-light noise alike, are computationally prohibitive and insufficiently interpretable. We therefore propose a fast, deterministic, denoising software pipeline specifically tailored to vocal fold motion, made available via GitHub. First, the low-light noise is reduced using total variation (TV-) denoising along the time axis. This reduces the noise level suitable for a classical spatial low-pass filter to remove the honeycomb artifacts, with subsequent histogram equalization to restore contrast. The method was evaluated on a proof-of-concept data set of 25 nasal HSV recordings, recorded at 10,000 Hz. The performance of the method was first evaluated with respect to consistency with the synchronously recorded, noise-free audio, by calculating the harmonics-to-noise ratio (HNR), based on the audio-derived fundamental frequency (f_o). For each video, the proposed pipeline significantly increased HNR (mean 8.1 dB), showing that the f_o-periodic vocal fold motion becomes more pronounced after applying the denoising pipeline. Secondly, the overall visual quality improvement was evaluated using the Naturalness Image Quality Evaluator (NIQE). Applying TV denoising with subsequent filtering decreased the NIQE scores (an increase in image quality) significantly, when compared to the unprocessed videos (mean decrease of 67.5%). Filtering without prior TV increased the visual quality of the unprocessed videos significantly less, with a mean NIQE score reduction of only 61.6%, TV denoising is therefore an effective preprocessing method. The computations were performed on consumer-grade hardware, showing that the proposed method is a fast and understandable denoising software solution, which improves visual quality measurably, while remaining faithful to the true vocal fold motion, therefore improving subsequent analysis.