DRome: A Deep Learning-Based Mobile Vision System for Real-Time Range of Motion Evaluation
Ankita Mukherjee, Helena Alvarez, Xiaorong Zhang, Zhuwei Qin, Charmayne Mary Lee Hughes · 2024
Musculoskeletal disorders (MSDs) are highly prevalent in modern workplaces, leading to physical pain, reduced mobility, reduced work productivity, and lost time due to sickness. In the physical rehabilitation context, the measurement of joint angle and range of motion (ROM) is a crucial component of musculoskeletal assessment and provides the clinician with valuable information about the flexibility and function of a particular joint, which is used to plan MSD rehabilitation protocols and monitor the effectiveness of therapies. The recent phenomenal development in deep learning-based computer vision and mobile computing have provided a means to track ROM from images captured by a single mobile camera, thereby allowing healthcare providers to assess progress without the need for in-person evaluation. In this paper, we propose DRome, a real-time mobile vision system that leverages deep learning for the evaluation of joint angle and ROM. Initial investigations indicate that the deep learning-based MoveNet model could not accurately measure joint angle (RMSE = 12.02). Subsequently, we assessed whether enhanced performance could be achieved through the utilization of machine learning models, with findings revealing that distinct motions necessitated the use of distinct, optimized machine learning algorithms (RMSE range = 3.70 - 7.29). After demonstrating that the DRome system yielded acceptable performance for the clinical setting, the system was developed for the Android OS, with testing indicating a real-time processing speed of 11–16 frames per second and less than 25 MB memory utilization. While exploratory, this work could lead to a paradigm shift in the way that MSD rehabilitation is managed, especially for individuals who reside in medically underserved areas who have reduced access to quality healthcare.