SAT0642 A cad system in hep-2 iif reading: a multicentre study

Background The indirect immunofluorescence (IIF) on HEp-2 cells is the recommended technique for anti-nuclear antibodies (ANA) detection. Automation of IIF image reading can provide a reliable basis for cost-effective serological diagnostics. Computer-aided diagnosis (CAD) tools based on digital imaging reading can help us to overcome the reader subjectivity. In a recent work 1 we assessed the inter-observer variability using digital images showing the 74.1% of agreement. It improved using “ground truth” as gold standard. Objectives To compare classification accuracy between readings provided by expert readers belonging to three different laboratories and those automatically returned by CAD. Methods We acquired 1771 images from 583 consecutive sample with an acquisition unit using HEp-2 cells (MBL) at 1:80 screening dilution. Each image was blindly classified as positive, negative, or weak positive by two experienced physicians, with more than ten years of experience in IIF, for each of the three laboratories. We obtained gold standard on the basis of annotations provided by expert physicians. We described a CAD system for HEp-2 classification that relies upon features provided by a deep neural network architecture, namely an Invariant Scattering Convolutional Network (Scatnet). We therefore compared human readings with automatic classification provided by a CAD. Results The dataset contains 215 positive samples, 136 weak positive, 219 negative. The CAD-system classification and experts showed a sensitivity of 93% vs 92.9%, 79.4% vs 74.5%, and 92.2% vs. 92.8% on positive, negative and weak positive, respectively (Fig 1). The CAD-system obtained an accuracy of 89.5%, slightly better than average experts’ classification (88.5%) and, interestingly, it better recognised the weak positive samples. Conclusions Solid gold standard is essential for use CAD systems in routine work lab. The CAD-system classification and the 3 experts have provided comparable results. Laboratories’ agreement improves using digital images and comparing each single human evaluation to a potential reference data and for this reason nowadays the CAD system should be considered a reliable tool of standardisation reducing the inter-laboratory variability. Reference [1] Rigon A, Infantino M, et al. Autoimmun Rev2017. Disclosure of Interest None declared