A NOVEL SUPERVISED METHOD FOR CORNEAL FIBRES TORTUOSITY CLASSIFICATION USING A MULTI-SCALE-MULTI-WINDOW APPROACH

We present a novel method for classifying corneal nerves images in 4 levels of tortuosity. We use multi-scale features and a new multi-window algorithm for curvature estimation. A wrapper using logistic ordinal regression or an embedded method based on random forests is applied for feature and scale selection. Tested with 90 images classified by 3 clinicians, the accuracy of our system is 88.89%, 76.67%, 77.22 % against each observer, outperforming the best one hold out (76.67%,76.67%,75%). Clinical Motivation I Corneal fibres appear QUALITATIVELY more tortuous in some diseases (e.g. keratoconic corneas [1, 2], dry eyes [3]); I Tortuosity significantly higher in Diabetic Retinopathy (DR), one of the leading pathological causes of blindness among working-age people in developed countries [4]. Technical Motivation MAIN ISSUE: QUALITATIVE assessment leads to high inter-observer and intra-observer variability. GOAL: to obtain an objective and reproducible tortuosity evaluation. Q1. Does scale play a role? Q2. Curvature seems a good feature: how accurately can it be estimated? Q3. Previous work: is this the right approach? Proposed Approaches IDEA: tortuosity is likely to be based on ophthalmologist’s experience and hardly captured by a formula- hence classifier: A Multiple-Window Approach for Digital Curvature Estimation 1. At each pixel, apply local ellipse fitting and line fitting using the smallest window size in a pre-defined range R = [