An automatic AVR biomarker assessment system in retinal imaging

Retinal Imaging, a non-invasive way to scan the back of the eye, provides a mean to extract different possible biomarkers, such as Artery and Vein Ratio (AVR). AVR is a well-known biomarker for various diseases, such as diabetes, glaucoma, hypertension, etc. The main objective of this paper to propose a fully automatic method to measure the AVR. The research hypothesis is that the system generated AVR is not significantly different from the ground truth. We have tested the system performance on publicly available INSPIRE-AVR (Iowa Normative Set for Processing Images of the REtina) dataset which contains 40 high-resolution colour fundus camera images and an AVR reference standard. The prerequisite for AVR measurement is the classification of retinal vessels (into arteries and veins) and the estimation of the vessel width. The images were classified into arteries and veins using Locally Consistent Gaussian Mixture Model (LCGMM) unsupervised classifier. The vessel width was estimated using the proposed Wavelet transform method from pre-processed images. Images pre-processing was performed using homomorphic filtering. Obtained results are compared with the vessel width calculated using the most common canny edge detector method. The calculated AVR was evaluated using two methods namely- Knudtson and Goatman, by utilizing the calculated vessel’s widths. The system-generated AVR results were compared with the ground truth (manually annotated by observer 1 (Ob1)), and statistical analysis was performed using a Student’s t-test. Furthermore, the validation of system-generated AVR values with respect to (w.r.t) the ground truth was done by utilizing a Bland–Altman (BA) plot. Student’s t-test shows no significant difference in the AVR measured using Knudtson blue(p-value is 0.805 > 0.05) and Goatman (p-value is 0.652 > 0.05) methods w.r.t Observer 1 (Ob1) when vessel width was measured using Wavelet transform. However, there was a significant difference between the AVRs by Ob1 and the system (with Knudtson: p-value is 0.01 < 0.05 and with Goatman: p-value is 0.02 < 0.05) when the vessel width was measured using the Canny edge detector. Bland Altman’s analysis shows that both the Ob1 and the system (with width calculated using Wavelet method and the AVR calculated using Knudtson and Goatman formula) have no substantial bias in AVR estimation. Furthermore, the observed bias between the AVR measurements was very low at 0.003. Further, from BA plot it has been seen that the limits of agreement for the system where width was obtained using canny was much wider as compared to the system when the wavelet transform was used to calculate the width. Further, our system generated average accuracy of 99.7% and 99.5% using Kundson and Goatman formula w.r.t Ob1. and outperform the existing method.