Automatic classification of serrated patterns in direct immunouorescence images

Abstract

Direct immunofluorescence (DIF) images are used by clinical experts for the diagnosis of autoimmune blistering diseases. The analysis of serration patterns in DIF images concerns two types of patterns, namely n- and u-serrated. Manual analysis is time-consuming and challenging due to noise. We propose an algorithm for the automatic classification of serrated patterns in DIF images. We first segment the epidermal basement membrane zone (BMZ) where n- and u-serrated patterns are typically found. Then, we apply a bank of B-COSFIRE filters to detect ridges and determine their orientations with respect to the BMZ. Finally, we classify an image by comparing its normalized histogram of relative orientations with those of the training images using a nearest neighbor approach. We achieve a recognition rate of 84.4% on a UMCG data set of 416 DIF images, which is comparable to 83.4% by clinical experts.