Mucociliary clearance is an essential defence mechanism against chronic airway infection and inflammation. Defects in ciliary motility are either primary, as in primary ciliary dyskinesia (PCD), or secondary. Identification of mucociliary clearance defects allows the implementation of appropriate management. High-speed video-microscopy (HSVM) is used to assess cilia motility from nasal biopsy samples. It is a time consuming and subjective requiring significant expertise. Computer vision can improve the identification of cilia motility defects by minimising subjectivity and reducing the cost and time to analyse samples.

Using an artificial intelligence platform (Intel® Geti?), we have trained several models using archived HSVM videos from patients referred to the Royal Brompton Hospital who were diagnosed with PCD and display a range of ciliary motility phenotypes and non-PCD controls. The videos used are converted to optical flow to provide temporal information to the machine learning algorithm.

We are training the platform to classify different categories of beat pattern: Immotile, Normal, Reduced Amplitude and Rotation. Models also include assessing sample quality and cilia beating orientation. The preliminary data based on projects currently in development are promising: the model classifying normal beating vs immotile cilia (around 30,000 frames) has a predictive accuracy of 100% and the beat pattern recognition model (around 25,000 frames) has a predictive accuracy of 97%.

Further training and testing are ongoing, and more models are being developed to include a greater range of motility phenotypes and to encompass chronic inflammatory lung diseases.