Abstract / Summary
Background/Aim: Accurate and consistent interpretation of HEp-2 cell patterns, considered the gold standard in autoimmune disease diagnosis, remains challenging due to inter-expert variability. This study aims to develop a hybrid deep learning architecture for the automated and highly accurate classification of HEp-2 cell images.Methods: The proposed approach integrates DenseNet-121, which excels at capturing local morphological features, with Vision Transformer (ViT-B16), known for modeling global contextual relationships. These models are combined using a weighted ensemble strategy. The performance of the model was evaluated on the MIVIA (ICPR 2012) dataset under a strict source-image split protocol, which explicitly prevents the object-level data leakage frequently overlooked in the literature.Results: The proposed hybrid model achieved an overall accuracy of 96.01% under these rigorous test conditions, without the use of synthetic data augmentation. A key methodological contribution is its ability to effectively distinguish between Homogeneous and Fine Speckled patterns, which are frequently misclassified by experts. Notably, the model achieved 100% precision for the Fine Speckled class. Additionally, it was empirically demonstrated that a small batch size (Batch Size = 8) improves generalization performance in fine-textured biomedical images.Conclusion: These findings demonstrate the robust feature extraction capabilities of the proposed hybrid architecture, establishing a reliable, realistic, and methodologically transparent baseline for future computer-aided diagnosis (CAD) systems.