Classification of Anti-nuclear Antibody Pattern Using Deep Learning

Autoimmune diseases are characterized by the situation when the immune system attack body’s own cells. The anti-nuclear antibodies (ANA) serve as biomarkers for diagnosis and identification of such diseases. This work presents a novel deep learning-based pipeline for the hierarchical classification of ANA patterns from HEp-2 cell images, addressing the challenges of limited data and overfitting prevalent in medical image analysis. The Hep-2 classification tree consists of 3 levels of classification with a total of 29 classes. The proposed approach introduces a Siamese network—based architecture, further enhanced by integrating ResNet-50 as a feature extractor to leverage transfer learning to improve classification performance. To overcome data scarcity, the Neural Style Transfer is applied to generate high-quality synthetic images along with other data augmentation methods to significantly expand the training set and improve model generalization. Among three levels of Hep-2 classification tree, 87–96% accuracy is achieved for level 1 and level 2 classes using the ResNet101—based classification model. For level 3, having the minimum number of data per class, an ensemble approach combining two ResNet-50-based Siamese networks with varied dropout rates is proposed. This has consistently achieved 90–97% accuracy across seven Level 3 ANA subclasses. Experimental Analysis shows that the proposed approach outperforms traditional CNN- and SVM-based methods, especially in low-data scenarios. This work not only advances automated ANA pattern analysis but also lays the base for developing comprehensive end-to-end diagnostic systems, with promising results for clinical practice and future research in medical image classification.