Adaptive fusion of swin transformer hierarchical features and handcrafted texture descriptors for ANA pattern classification

Background Accurate identification of human epithelial type-2 (HEp-2) cell images is significant for diagnosing autoimmune diseases. Objective This study aims to improve classification accuracy on the HEp-2 image dataset by presenting a robust hybrid model that integrates deep hierarchical representations extracted using the Swin Transformer architecture with traditional handcrafted texture features. Methods The Swin Transformer efficiently captures both fine-grained and global spatial information. In parallel, texture-based descriptors such as Local Binary Patterns (LBP), Gabor filters, and the Gray-Level Run Length Matrix (GLRLM) are utilized to extract detailed micro-textural variations in cell structures. These complementary features are fused using an adaptive method with an attention-driven weighting mechanism, permitting optimal integration of heterogeneous information. Results Experimental results on a publicly available HEp-2 benchmark dataset establish that the proposed technique outperforms numerous conventional models, accomplishing a peak accuracy of 98.01%, along with excellent precision, recall, and F1-score across all types. Additionally, ablation studies validate the individual contributions of each component within the architecture. Conclusion Overall, the proposed technique displays strong potential for reliable HEp-2 cell classification in medical diagnostic applications.