Comparison of artificial intelligence applications and commercial system performances using selected ANA IIF images

Accurate and accessible classification of anti-nuclear antibodies (ANA) through indirect immunofluorescence (IIF) imaging is crucial for diagnosing autoimmune diseases. However, many laboratories, particularly those with limited resources, lack access to expensive commercial systems for automated analysis. This study evaluates the performance of an application developed by expert physicians using an artificial intelligence application (Microsoft Azure) to classify ANA IIF images. The results are compared with EuroPattern to assess the potential of AI in assisting laboratory experts, especially in resource-limited settings. A total of 648 ANA IIF images from the EuroPattern archive were used to train an AI model across nine classes with varying fluorescence intensities (+ to + + + +). Testing was conducted with 96 images, ensuring clarity by excluding mixed patterns. Microsoft Azure's Custom Vision service was employed for labeling and prediction. Expert evaluations, EuroPattern results, and AI classifications were compared. Both EuroPattern and the Azure-based AI model achieved 100% sensitivity, specificity, and accuracy in positive and negative discrimination. EuroPattern had an intraclass correlation coefficient (ICC) of 0.979, and the Azure-based AI model had an ICC of 0.948, indicating slightly lower performance. EuroPattern outperformed the Azure-based AI model in recognizing homogeneous, speckled, centromere, and dense fine-speckled patterns. The Azure-based AI model performed better in identifying cytoplasmic reticular/AMA-like patterns. The results suggest that AI-based image analysis tools, such as Azure, can be valuable for diagnostics in resource-limited labs. However, further testing with larger datasets and routine patient samples is needed to confirm their effectiveness in real-world clinical settings.