Comparison of the Capacity of Several Machine Learning Tools to Assist Immunofluorescence-Based Detection of Anti-Neutrophil Cytoplasmic Antibodies.

Abstract

The success of artificial intelligence and machine learning is an incentive to develop new algorithms to increase the rapidity and reliability of medical diagnosis. Here we compared different strategies aimed at processing microscope images used to detect anti-neutrophil cytoplasmic antibodies, an important vasculitis marker: (i) basic classifier methods (logistic regression, k-nearest neighbors and decision tree) were used to process custom-made indices derived from immunofluorescence images yielded by 137 sera. (ii) These methods were combined with dimensional reduction to analyze 1733 individual cell images. (iii) More complex models based on neural networks were used to analyze the same dataset. The efficiency of discriminating between positive and negative samples and different fluorescence patterns was quantified with Rand-type accuracy index, kappa index and ROC curve. It is concluded that basic models trained on a limited dataset allowed for positive/negative discrimination with an efficiency comparable to that obtained by conventional analysis performed by humans (0.84 kappa score). More extensive datasets and more sophisticated models may be required for efficient discrimination between fluorescence patterns generated by different auto-antibody species.