Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Signal processing and machine learning techniques for Doppler ultrasound haemodynamic measurements

Latham, Jack 2020. Signal processing and machine learning techniques for Doppler ultrasound haemodynamic measurements. PhD Thesis, Cardiff University.
Item availability restricted.

[thumbnail of PhD Thesis]
PDF (PhD Thesis) - Accepted Post-Print Version
Download (35MB) | Preview
[thumbnail of Cardiff University Electronic Publication Form] PDF (Cardiff University Electronic Publication Form) - Supplemental Material
Restricted to Repository staff only

Download (328kB)


Haemodynamic monitoring is an invaluable tool for evaluating, diagnosing and treating the cardiovascular system, and is an integral component of intensive care units, obstetrics wards and other medical units. Doppler ultrasound provides a non-invasive, cost-effective and fast means of haemodynamic monitoring, which traditionally necessitates highly invasive methods such as Pulmonary artery catheter or transoesophageal echocardiography. However, Doppler ultrasound scan acquisition requires a highly experienced operator and can be very challenging. Machine learning solutions that quantify and guide the scanning process in an automatic and intelligent manner could overcome these limitations and lead to routine monitoring. Development of such methods is the primary goal of the presented work. In response to this goal, this thesis proposes a suite of signal processing and machine learning techniques. Among these is a new and real-time method of maximum frequency envelope estimation. This method, which is based on image-processing techniques and is highly adaptive to varying signal quality, was developed to facilitate automatic and consistent extraction of features from Doppler ultrasound measurements. Through a thorough evaluation, this method was demonstrated to be accurate and more stable than alternative state-of-art methods. Two novel real-time methods of beat segmentation, which operate using the maximum frequency envelope, were developed to enable systematic feature extraction from individual cardiac cycles. These methods do not require any additional hardware, such as an electrocardiogram machine, and are fully automatic, real-time and highly resilient to noise. These qualities are not available in existing methods. Extensive evaluation demonstrated the methods to be highly successful. A host of machine learning solutions were analysed, designed and evaluated. This led to a set of novel features being proposed for Doppler ultrasound analysis. In addition, a state of- the-art image recognition classification method, hitherto undocumented for Doppler ultrasound analysis, was shown to be superior to more traditional modelling approaches. These contributions facilitated the design of two innovative types of feedback. To reflect beneficial probe movements, which are otherwise difficult to distinguish, a regression model to quantitatively score ultrasound measurements was proposed. This feedback was shown to be highly correlated with an ideal response. The second type of feedback explicitly predicted beneficial probe movements. This was achieved using classification models with up to five categories, giving a more challenging scenario than those addressed in prior disease classification work. Evaluation of these, for the first time, demonstrated that Doppler scan information can be used to automatically indicate probe position. Overall, the presented work includes significant contributions for Doppler ultrasound analysis, it proposes valuable new machine learning techniques, and with continued work, could lead to solutions that unlock the full potential of Doppler ultrasound haemodynamic monitoring.

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Engineering
Uncontrolled Keywords: Doppler Ultrasound; Envelope Estimation; Beat Segmentation; Features; Machine Learning; Classification.
Date of First Compliant Deposit: 9 March 2021
Last Modified: 08 Mar 2022 02:30

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics