Learning the Relationship Between Lung Ultrasound Appearance and Pulmonary Oedema Clinical Score
Principal Supervisor: Prof. Alison Noble (Institute of Biomedical Engineering, Oxford)
Co-Supervisors: Prof. Najib Rahman (Oxford), Dr. Rob Janiczek (GlaxoSmithKline)
There is a clinical need for an imaging technique that non-invasively measures pulmonary oedema for managing patient care and for use within clinical trials. Traditional techniques such as x-ray are often sufficient for diagnosing oedema but not for assessing severity or measuring the response to interventions. Lung ultrasound (LUS) has recently gained traction as a technique for assessing pulmonary oedema. LUS offers advantages over traditional assessments such as x-ray, PET, and MRI due to its ease of access and ability to image acute patients at the bedside (e.g. A&E, ICU). However, LUS suffers from several of the common limitations experienced by ultrasound techniques, namely inter-observer variability and the need for experts to interpret scans well.
In healthy lungs, ultrasound waves do not penetrate the air/tissue interface therefore lung parenchyma appears dark on LUS scans. In the presence of pulmonary oedema, however, small reverberations at the edge of the lung manifest as comet tail artefacts (called b-lines) that appear to traverse across the lung. The number of b-lines have been shown to correlate with the severity of pulmonary oedema as measured by alternative techniques (e.g. indicator dilution) and with clinical outcomes. LUS is currently scored by manually counting the number of b-lines that are present in various regions of the lung. This becomes difficult in regions with severe pulmonary oedema as the comet tail artefacts become confluent and the differentiation of b-lines becomes subjective.
This studentship will develop an image analysis algorithm to automatically score LUS (e.g. using machine learning to learn the relationship between image appearance and a clinical score). The project will utilize data from a recently completed clinical study at Oxford (https://clinicaltrials.gov/ct2/show/NCT01949402). In the study, subjects with chronic obstructive pulmonary disease; interstitial lung disease; and patients on hemodialysis (to replicate acute pulmonary oedema/heart failure) were scanned using LUS and scored as described above by an expert observer. CT was additionally collected in subjects to serve as a comparator. A subset of data of these subjects will be used to train the algorithm with either LUS scores or CT attenuation density serving as ground truth. Results from the remainder of the subjects (not used for training) will then be utilized to test the algorithm against the expert reader and/or CT density.
G. Volpicelli, V. Caramello, L. Cardinale, A. Mussa, F. Bar, and M. F. Frascisco, “Bedside ultrasound of the lung for the monitoring of acute decompensated heart failure,” Am. J. Emerg. Med., vol. 26, no. 5, pp. 585–591, 2008.