Fine-grained and multi-pattern anti-nuclear antibody recognition: A new dataset and framework.

Abstract

Indirect immunofluorescence (IIF) staining imaging on human epithelial cell (HEp-2) is a standard technique for anti-nuclear antibodies (ANA) screening. The application of deep neural networks for automated IIF image processing has led to remarkable progress in ANA pattern recognition. However, existing research has predominantly focused on coarse-grained and single-pattern recognition, despite the fact that the International Consensus on Antinuclear Antibody Patterns (ICAP) has defined 29 fine-grained patterns and IIF images often present multiple concurrent ANA patterns. In this study, we contribute a large scale IIF image dataset for fine-grained and multi-pattern recognition, which comprises up to 40 thousand high-resolution images and precise annotations of 17 ICAP recommended basic patterns and 5 pattern combinations from technologists with >10 years ANA experience. In addition, we introduce a neural network based recognition framework for this task, which consists of a mitotic detection network and a pattern classification network. To enhance the effectiveness and robustness of the recognition framework, we employ multiple strategies, including dataset resampling, automatic image augmentation and prediction aggregation. Our model achieved an F-score of 0.883 on the test dataset, significantly outperforming human technologists. Furthermore, with the assistance of the developed model, we observed a significant and consistent improvement in human performance. This approach enables scalable, automated and multi-pattern ANA reading aligned with ICAP standards.