A novel image-based algorithm to support future remote assessment of chewing function

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Chewing difficulty is associated with poorer physical and mental health. Objective measurement of chewing function is currently limited to methods that require specialist lab equipment (for example lab-based manipulation or comminution tests). A remote method for estimating chewing-related metrics would support the development of accessible and scalable means of detecting and monitoring chewing-related health issues. Method This paper details initial work on developing a novel smartphone-based, image-analysis algorithm to estimate particle-size metrics from photographs of masticated raw carrot deposited in a Petri dish. A two-stage image-processing pipeline was developed, comprising geometric calibration and particle segmentation, enabling estimation of particle-size metrics from smartphone photographs. The algorithm was implemented both as a Python script and as a graphical user interface (GUI), the latter allowing semi-automated per-image adjustment of calibration and segmentation parameters with visual feedback. The script and graphical user interface (GUI) are publicly available: osf.io/kgx9f Results Performance was evaluated on artificially generated test images to benchmark the algorithm against ground-truth data with known sizes, while also assessing data-collection processes, usability and key challenges. On the more realistic synthetic-particle images, which include overlapping particles of varying size, the algorithm produced a cumulative area error of 20.6% and a mean per-particle relative error of 21.1% (median 14.8%; mean absolute error 1.02 mm 2 ), with the largest errors occurring for the smallest and most overlapped particles. These synthetic-particle figures reflect performance on successfully matched particles; end-to-end performance including undetected particles would be lower. On simpler geometric-shape images, total-area relative error was lower, at 3.10% for the large-shape set, 3.00% for the small-shape set, and 3.12% across all shapes. Conclusion The algorithm demonstrates quantifiable analytical performance in estimating chewing-related particle-size metrics from smartphone images, supporting its potential use as a foundation for future remote measurement of chewing function. However, the method is not yet clinically validated. In its current form, the method achieves its best performance through a semi-automated, GUI-based workflow, which introduces a subjective element but provides a practical way to handle heterogeneous image conditions and inform future automation.