| Summary: | This thesis introduces several novel, noninvasive lung function assessment approaches in which we incorporate computer vision techniques to remotely compute standard clinical Pulmonary Function Testing (PFT) measures. Using a single depth sensor, a dynamic 3-D model of a subject's chest is reconstructed and used to generate chest volume-time data by estimating the chest volume variation throughout a sequence. Following computation of multiple keypoints and calibration of volume-time data to present real volume of exchanged air, 7 Forced Vital Capacity (FVC) measures and 4 Slow Vital Capacity (SVC) measures are computed. Evaluation on a dataset of 85 patients (529 sequences), attending a respiratory outpatient service for spirometry, shows a high correlation between the proposed depth-based PFT measures and the measures from a spirometer. Trunk motion during PFT affects the accuracy of these results, so the natural reaction of the subject's body to maximal inhalation and exhalation, must be decoupled from the chest-surface breathing motion. We present an automatic, open source data acquisition and calibration pipeline in which two opposing depth sensors are calibrated and used to reconstruct a well-defined dynamic 3-D model of the trunk during PFT performance. Our proposed method is able to reconstruct dynamic 3-D models with accurate temporal frame synchronisation and spatial registration. Then, we propose a whole body depth-based photoplethysmography (dPPG) approach which allows subjects to perform PFT, as in routine spirometry, without restraining their natural trunk reactions. By decoupling the trunk movement and the chest-surface respiratory motion, dPPG obtains more accurate respiratory volume-time data which improves the accuracy of the estimated PFT measures. A dataset spanning 35 subjects (298 sequences) was collected and used to illustrate the superiority of the proposed dPPG method by comparing its measures to those provided by a spirometer and the single Kinect approach. Although dPPG is able to improve the PFT measures accuracy to a significant extent, it is not able to filter complex trunk motions, particularly at the deep forced inhalation-exhalation stage. To effectively correct trunk motion artifacts further, we propose an active trunk shape modelling approach by which the respiratory volume-time data is computed by performing principal component analysis on temporal 3-D geometrical features, extracted from the chest and posterior shape models in R3 space. We validate the method's accuracy at the signal level by computing several comparative metrics between the depth-based and spirometer volume-time data. Evaluating on the dPPG PFT dataset (300 PFT sequences), our trunk shape modelling approach outperforms the single Kinect and dPPG methods.
|
|---|
