Uncertainty analysis of clinically relevant distances derived using photogrammetric intersection for the assessment of motor-speech-control in children

Abstract. Perceptual analysis is the current benchmark standard used by speech-language pathologists in diagnosing speech sound disorders (SSDs). Yet, related research indicates access to objective measures could improve the assessment process. Recent technological advances have contributed to developing AI-based methods to provide clinicians access to objective measures for speech-motor control by calculating inter-landmark distances of anatomically relevant facial landmarks. However, landmarks placed by AI-based methods extract the landmarks’ coordinates without associated uncertainties. Consequently, inter-landmark distances extracted as objective measurements also lack uncertainty information, potentially compromising their suitability for assessment purposes. In contrast, photogrammetry can predict facial inter-landmark distances and their uncertainties through intersection and variance propagation. In this paper, we use a combination of the markerless BlazeFace algorithm and photogrammetry to examine how different weightings of the image observations, introduced for the photogrammetric intersection, impact the assessment of whether the calculated inter-landmark distances significantly change during the production of spoken words. We selected 16 inter-landmark distances to assess jaw movement. We analysed the movements of five children saying 10 words. Overall, four different weightings and two different camera setups were tested. Setup 1 used 2 cameras, and setup 2 used 3 cameras. The weightings based on comparing the BlazeFace landmarks to a reference were too large when applied to setup 1. They did not allow the reliable determination of inter-landmark distance changes as predicted by current literature depending on the camera setup used. Smaller weights were able to be statistically tested for jaw movements correctly. For setup 2, all weights could detect inter-landmark distances reliably.