fetching data ... 
Background: Artificial Intelligence (AI) is increasingly revolutionizing medicine by creating intelligent systems that enhance diagnostics and treatment. In rheumatology, the gold standard for studying autoantibodies in systemic autoimmune diseases is Antinuclear Antibody (ANA) testing via Indirect Immunofluorescence (IIF). However, ANA interpretation is subjective and depends on observer expertise. AI, particularly deep learning, offers solutions by accurately recognizing ANA patterns, reducing reporting errors, minimizing inter-observer variability and predicting clinically relevant autoantibodies. Research demonstrates AI’s high agreement with expert interpretations, highlighting its potential to streamline rheumatological diagnostics. Deep learning models like YOLOv8, DenseNet121, and ResNet50V2 have shown immense potential in automating ANA image classification. In our study, we tested and validated an AI-based machine learning algorithm to identify common ANA patterns from our dataset at a tertiary care rheumatology center, including homogenous, dense fine nuclear speckled, centromere, and fine and coarse speckled ANA pattern images.
To use various CNN (Computational Neuronal Networks) models for the interpretation of ANA IIF patterns.
To test and validate the best-performing algorithm.
Methods: The study was a prospective, observational study conducted over 18 months. A total of 803 samples were collected, processed and tested using Indirect Immunofluorescence with Hep-2 cell coated slides. Prepared samples were categorized into ICAP (International Consensus on ANA Patterns) standardized patterns (AC-1 to AC-5) and negative samples, excluding any mixed or non-consensual patterns. The ANA patterns were independently evaluated by three expert observers. The AI model development used various convolutional neuronal networks (CNN). The CNNs included were YOLOv8, DenseNet121, InceptionResNetV2, InceptionV3 MobileNetV2, ResNet50V2, VGG19 and Xception. The CNNs were trained on a 90:10 dataset split for training and validation, with a 5-fold cross-validation to prevent overfitting. Performance was evaluated using confusion matrices, ROC curves, precision, recall, and F1 scores, targeting six ANA classes in a multi-label classification framework. Image data were acquired using the Image Navigator system at 1392x1040 pixels and classified into ICAP categories through a multi-level approach. Preprocessing involved converting RGB images to black-and-white by removing non-essential channels and normalizing intensities. The final model validation was performed on novel, unseen ANA samples to ensure robustness and generalizability.
Results: YOLOv8 consistently outperformed other models. The overall accuracy of YOLOv8-Base model was 79.06%, while accuracy of the YOLOv8-Pretrained model was significantly better with 89.36% accuracy. For multi-class classification, YOLOv8 pretrained model showed superior precision, F1-score, and class-wise results. In case of binary tasks the YOLOv8 pretrained model showed accuracies of 97.52% for negative vs non-negative, 97.4% for metaphase vs non-metaphase, 92.71% for AC-1 to AC-3, and 93.52% for AC-4 vs AC-5 classifications. Other models like DenseNet121, InceptionResNetV2 and InceptionV3 struggled with accuracy. MobileNetV2 and Xception models showed the weakest performance.
Conclusion: This study demonstrates the power of deep learning for ANA pattern classification, particularly using advanced models like YOLO-v8. The hierarchical multi-class approach offers a robust solution to the complexity inherent in ANA diagnostics, showing superior performance compared to many previous studies. Additional research should focus on refining the model’s ability to distinguish between visually similar ANA patterns, such as AC-4 and AC-5. This could involve the use of more sophisticated feature extraction techniques or the integration of external clinical data to improve pattern differentiation. Furthermore, AI can facilitate the identification of mixed patterns, which may present unique challenges in diagnosis.
REFERENCES: NIL.
Acknowledgements: NIL.
Disclosure of Interests: None declared.
© The Authors 2025. This abstract is an open access article published in Annals of Rheumatic Diseases under the CC BY-NC-ND license (