Artificial intelligence for automated classification of antinuclear-antibody indirect immunofluorescence patterns

Objectives: This study aims to interpret antinuclear-antibody indirect immunofluorescence (ANA-IIF) patterns using a multistage deep learning (DL) classification framework. Methods: A prospective, observational investigation was conducted over 18 months, aiming to develop and validate algorithms for classifying ANA-IIF patterns (N = 803), focusing on positive samples with International Consensus on Antinuclear Antibody Patterns standard patterns AC-1 to AC-5 and negative samples, while excluding mixed or nonconsensual patterns. Eight convolutional neural network (CNN) architectures were benchmarked against a multistage DL classification framework, where it first classifies an image into positive vs negative patterns, then metaphase stained vs not stained, and finally subclassifications within the stained and not-stained groups. All models were trained and evaluated using a 90:10 train-test split with 5-fold cross validation. Results: The multistage YOLOv8 pretrained models showed strong performance at each stage, achieving accuracies of 98.02% ± 0.49%, 96.57% ± 0.57%, 87.22% ± 1.11% (stained group), and 92.35% ± 1.5% (not-stained group), with an overall accuracy of 92.27 ± 1.01. In comparison, the multiclass YOLOv8 pretrained model achieved an overall accuracy of 87.16% ± 1.0%. The multistage YOLOv8 model consistently outperformed the other CNN architectures, reaching an average accuracy difference of 45.86 ± 1.25. Conclusions: This study demonstrates the advantage of multistage physician-centric design and implementation of a DL model for antinuclear antibody (ANA) pattern classification, particularly using advanced models like YOLO-v8. The hierarchical multistage approach offers a robust solution to the complexity inherent in ANA diagnostics, showing superior performance compared to other CNN architectures and a multiclass single model/classifier.