Automatic Classification of Antinuclear Antibody Patterns With Machine Learning

Abstract

Antinuclear antibodies (ANA) are important diagnostic markers in many autoimmune rheumatological diseases. The indirect immunofluorescence assay applied on human epithelial cells generates images that are used in the detection of ANA. The classification of these images for different ANA patterns requires human experts. It is time-consuming and subjective as different experts may label the same image differently. Therefore, there is an interest in machine learning-based automatic classification of ANA patterns. In our study, to build an application for the automatic classification of ANA patterns, we construct a dataset and learn a deep neural network with a transfer learning approach. We show that even in the existence of a limited number of labeled data, high accuracies can be achieved on the unseen test samples. Our study shows that deep learning-based software can be built for this task to save expert time.

Comparison of the Capacity of Several Machine Learning Tools to Assist Immunofluorescence-Based Detection of Anti-Neutrophil Cytoplasmic Antibodies

“…Indeed, neural networks may now be considered the gold standard for image analysis [ 12 ], and they are currently the basis of many current reports on ML classification for medical purposes [ 49 ]. However, while more and more powerful network architectures are continually being reported and tested [ 17 , 18 , 50 ], model setting and training quality are essential determinants of final performance. Available strategies will be rapidly listed below.…”

“…Another procedure facilitating the training of very complex models, dubbed transfer learning , consists of using a pretrained model and training only the outer layers to fit a specific dataset. This method permits the use of highly successful models trained on public image datasets such as ImageNet [ 54 ] with a reasonable computing load for ANA classification with Hep-2 cells [ 50 , 55 ].…”