Improving Primary Ciliary Dyskinesia diagnosis using Artificial Intelligence

Abstract

Improving Primary Ciliary Dyskinesia diagnosis using Artificial Intelligence: Primary ciliary dyskinesia (PCD) is an inherited autosomal-recessive disorder of motile cilia that results in chronic lung disease, rhinosinusitis, hearing impairment and subfertility. The estimated prevalence of PCD is ∼1 per 10,000 births, but it is more prevalent in populations where consanguinity is common. To diagnose PCD involves a combination of tests, in particular, electron microscopy (EM) essential for determining the type of ciliary ultrastructural defect. EM involves meticulous inspection of ~300 ciliary cross-sections, which is time-consuming and requires highly skilled and experienced diagnostic scientists. Machine learning offers an opportunity to improve diagnostic accuracy, reduce time to analyse samples, minimise the subjective element and significantly reduce costs.

NOUS, by COSMONiO, is the on-site deep-learning system that is being trained to detect the difference between healthy and abnormal cilia. Different teaching approaches are being tested that include (i) Detection of cilia that are usable for diagnosis (ii) Classification of normal and abnormal cilia.

Over 15,000 EM images of cilia have been input to NOUS and it is showing consistently improved accuracy as the training datasets enlarge. When testing it against diagnostic specialists (n=5) using blinded image datasets, an agreement of >75% in the classification of images was found. This is similar to the agreement measured between individual diagnostic specialists.

NOUS training is ongoing, and further datasets are being used to determine the sensitivity and specificity. Furthermore, it is currently being trialled alongside the current diagnostic test.