Recent neurological studies suggest that oculomotor alterations are one of the most important biomarkers to detect and characterize Parkinson’s disease (PD), even on asymptomatic stages. Nevertheless, only global and simplified gaze trajectories, obtained from tracking devices, are generally used to represent the complex eye dynamics. Besides, such acquisition procedures often require sophisticated calibration and invasive configuration schemes. This work introduces a novel approach that models very subtle ocular fixational movements, recorded with conventional cameras, as an imaging biomarker for PD assessment. For this purpose, a video acceleration magnification is performed to enhance small fixational patterns on standard gaze video recordings of test subjects. Subsequently, feature maps are derived from spatio-temporal video slices by means of convolutional layer responses of known pre-trained CNN architectures, allowing to describe the depicted oculomotor cues. The set of extracted CNN features are then efficiently coded by means of covariance matrices in order to train a support vector machine and perform an automated disease classification. Promising results were obtained through a leave-one-patient-out cross-validation scheme, showing a proper PD characterization from fixational eye motion patterns in ordinary sequences.
CITATION STYLE
Salazar, I., Pertuz, S., Contreras, W., & Martínez, F. (2019). Parkinsonian Ocular Fixation Patterns from Magnified Videos and CNN Features. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11896 LNCS, pp. 740–750). Springer. https://doi.org/10.1007/978-3-030-33904-3_70
Mendeley helps you to discover research relevant for your work.