4-year PhD Studentship: Advanced analytics applied to endoscopic analysis of upper airway function in horses.

Upper airway obstructions affect the health, welfare and performance of horses and are evaluated clinically by performing endoscopy. Numerous forms of pharyngeal and laryngeal obstructions occur; however, assessment of endoscopy recordings is subjective, with variable agreement between veterinarians (Ref:1,2) and no practical method to quantify the degree of airway obstruction. Subjective assessment by veterinarians suggests that resting endoscopy (widely performed in clinics and sales) is poorly predictive of the gold-standard exercising endoscopy (Ref:3). Furthermore, for many important and prevalent conditions the pathophysiology is poorly understood, limiting the efficacy of current treatment options.

A fully objective and automated method of analysis that can be applied to an endoscopic recording would be of great value 1) in providing a robust, objective means of characterising laryngeal movement; 2) to inform and benefit clinical decision making; 3) to inform pathophysiological clinical research; and 4) to improve clinical research intervention trials. Advances in artificial intelligence (AI) in the form of deep neural networks (DNN) provide new opportunities for dynamic measurement and tracking of different parts of the upper airway allowing for unbiased, objective, state-of-the-art assessment, and quantification of airway function. Our pilot study using DNNs, demonstrates that accurate segmentation of laryngeal structures is readily achievable.

Aims and Objectives

Aim: To apply advanced analytics, in the form of DNNs and other techniques, to endoscopy recordings of equine upper airway function. With the models and techniques developed undertake a series of clinically relevant studies to demonstrate the value of this state-of-the-art approach for both clinical and research purposes.

Objectives:

1. Develop objective, automated methods to identify and quantify selected upper airway structures on endoscopy images.
2. Determine the effect of laryngeal obstructions on rima glottidis diameter.
3. Establish whether caudal laryngeal retraction is a precursor to soft palate displacement.
4. Investigate the prediction accuracy of exercising laryngeal function from resting endoscopy.

Methodology

Advanced analytics is an umbrella term that includes (semi-)autonomous examination of data using sophisticated techniques and tools to reveal deeper insights, make predictions, or generate recommendations.

Developing an objective, automated method to identify and quantify upper airway structures will involve building and training a DNN that can identify and locate structures on an endoscopic recording, frame by frame (referred to as segmenting), such that they can be measured and quantified geometrically. Further DNNs or unsupervised learning approaches will be used to determine the effect of laryngeal obstructions on airway diameter by measuring rima glottidis size, demonstrating the effect of these conditions during the respiratory cycle and over an exercise period.

Furthermore, using similar approaches, it will be established whether an automated method can be used to understand better the pathophysiology of soft palate displacement. Establishing whether caudal retraction of the larynx can be identified as a precursor, and, if so, in what proportion of horses does caudal retraction occur, and over what period prior to displacement does it occur.

Finally, whether machine learning and predictive analytics applied to laryngeal function during resting endoscopy can better predict laryngeal function during exercise, than subjective assessment, will be established.

Keywords

Veterinary, equine, endoscopy, machine learning, artificial intelligence, deep neural networks

How to apply for this project

This project will be based in Bristol Veterinary School in the Faculty of Health Sciences at the University of Bristol.

Please visit the Faculty of Health Sciences website for details of how to apply