Deep learning for automated classification of antinuclear antibody patterns on HEp-2 indirect immunofluorescence images.

Standardizing the interpretation of ANA patterns is a persistent challenge in rheumatology due to inter-observer discordance. This research introduces an ensemble deep learning framework optimized for the International Consensus on ANA Patterns (ICAP) system. By integrating ResNet-50 and EfficientNet-B0 via validation-tuned weighting and logit-level averaging, we processed both a single-label benchmark and a high-variability clinical dataset containing overlapping (multi-label) patterns. Our model demonstrated high fidelity in the single-label setting (92.5% accuracy; MCC = 0.9193). Critically, in the independent hospital cohort, the system managed the complexities of co-occurring patterns with a 1.26% Hamming loss and a micro F1-score of 82.91%. By achieving "near-miss" accuracy (within two labels) in over 95% of clinical cases, this framework demonstrates its potential utility as a decision-support tool that may help mitigate the inherent variability associated with manual HEp-2 cell analysis.