fetching data ... 
Background To detect antinuclear antibodies (ANA) on the blood’s patients is a mostly used technique to diagnose immunologic diseases. The Indirect Immunofluorescence method (IIF) generates images with ANA patterns on them and is commonly used by immunology laboratories around the world. Those patterns are documented in the International Consensus of ANA Patterns (ICAP) and there are 29 of them at present, split into three main groups: Nuclear, Cytoplasmic and Mitotic. Classifying these patterns is a subjective task that has a strong dependence on the physician’s experience and training, thus the professionals need a second reader to reach a successful classification of images with these patterns. To support the professionals, we developed a Machine Learning model (ML), based on Convolutional Neural Networks (CNN) able to classify ANA patterns between positive and negative and then classifying only positive samples between the three main groups mentione.
Objectives Developing a machine learning model to help physicians reach a successful ANA pattern classification without a second reader and to train entry-level professionals to classify different ANA images by themselves.
Methods Using a public database with 2079 samples of ANA pattern images we defined the limitations of our model; we preprocessed the dataset and made a data augmentation process to avoid overfitting issues. To reach our objectives we define a model with two parts, the first one was a CNN classifying each sample between positive and negative, the second one was another CNN classifying the positive samples from the last step into the main three groups before mentioned. Following this idea, we proved 17 pre-trained CNN and compared their results based on commonly used metrics in machine learning: accuracy, precision, recall and F1. To prove our final model we train it with the mentioned augmented public dataset and test it with a private dataset with 445 samples from Inmuno21 Laboratory in Caracas, Venezuela.
Results Finally, our model reached acc: 93%, pre: 93%, recall: 93% and F1: 93% classifying between positives and negatives, improving the state of the art models presented before as a solution for this problem1. On the other hand, our model shows an acc: 75%, pre: 85%, recall: 92% and F1: 87% for the second stage, classifying positives samples in Nuclear, Mitotic and Cytoplasmic, becoming the first model documented that be able to classify ANA patterns images in this three groups with multilabel capability and positives-negatives previous discrimination.
Conclusion It’s necessary to make efforts to create a bigger datasets of ANA patterns images, better documented and with more samples for each label, since the major problem we found when we’ve developed the project was there was poor data and there was only one public dataset that had not samples for each label, the other public datasets were not available, had lower samples amount than the ones used or had no samples in the three main groups. Thus it was necessary to augment data to front the overfitting issues. Despite this, we made a model that shows high performance compared with state of the art works, and it can be used as a second reader or training partner for new professionals.
Reference [1]Cascio, D.; Taormina, V.; Raso, G. Deep Convolutional Neural Network for HEp-2 Fluorescence Intensity Classification. Appl. Sci. 2019, 9, 408.
Acknowledgements: NIL.
Disclosure of Interests None Declared.
Keywords: Diagnostic tests, Biomarkers, Artificial intelligence
DOI: 10.1136/annrheumdis-2023-eular.1781