Automated examination of relative afferent pupillary defect using pupil detection and tracking
Mathew, Melvin Julian
MetadataShow full item record
In traditional clinical practice, medical professionals typically examine ocular conditions by conducting a series of procedures using various medical equipment, and these examinations are subject to the environment in which the medical examination takes place. The variability in testing setup can potentially lead to inconsistency in medical assessments , however, supposing that these factors can be perfectly controlled, there is still a factor of inconsistency in the ocular condition assessment. Assessments are typically performed visually, which depends on the professional’s medical experience, knowledge, and personal health. More so, there is an increasing presence of ocular conditions globally , with a shortage and even stagnant presence of eye care professionals [3, 4]. Therefore, there is an increasing need for more consistent, accurate and automated approach to medical examination of ocular conditions. Swinging flashlight test is a medical procedure used to determine Relative Afferent Pupillary Defect (RAPD) , a common indicator of ocular disease. RAPD is the condition in which the pupillary light reflex response of both pupils differ to one another. The severity of RAPD is correlated to the similarity or dissimilarity of the pupillary light reflex responses . In this thesis, we propose automating RAPD examination using pupil detection and tracking algorithms to measure the pupil response during the swinging flashlight test. Specifically, we utilize image pre-processing techniques including temporal cropping, downsampling, and spatial cropping, to improve pupil detection and pupil size measurement by the Circular Hough Transform. RAPD severity score is obtained by performing Pearson Correlation Coefficient on the measured pupil responses, which quantifies the similarity or dissimilarity. We also perform post-processing techniques on the obtained pupil size measurements, including median filtering, thresholding, and moving average, to minimize noise and incorrect pupil detection measurements. The proposed algorithmic pipeline achieves 84% accuracy on the available testing set.