Automation for patient screening

Student thesis: Doctoral Thesis


Diagnostic imaging is the gold standard for differential diagnosis of disease, with ultrasound being the second most requested scan after X-ray with more than 8 million ultrasounds performed by NHS England in 2021 accounting for over 20% of all imaging performed. Ultrasound cross sectional imagery is used every day to make critical decisions that could drastically affect patient outcome. While diagnostic ultrasound cross sections are clearly defined within a clinical protocol, the clinician is solely responsible for acquisition and interpretation of ultrasound imagery, with few safeguards against human error. This Canon sponsored EngD looked at the potential of machine learning to standardise processes, reduce burden on users by automating the adherence to protocols, reduce the time required by streamlining workflows, and lower the skill requirement of the clinical user. The initial study was the first to characterise the response of neural networks for the classification of cross sections specified by the Japanese abdominal scanning protocol. This protocol, one of the largest ultrasound protocols ever studied, consists of 16 overlapping cross sectional views of the abdomen, and achieved a classification accuracy of 79.9%. This provided a baseline for a transfer learning study, utilising pre-trained neural networks to increase training efficiency and lead to an increase in accuracy to 83.9%. Small mobile networks were shown to be just as effective at classification of ultrasound at a fraction of the system resources, achieving comparable accuracies of 84.5%. Novel methods of cost reduction were explored to lower the burden of production of datasets for machine learning using power theory and active learning, providing a novel cost-effective framework for data collection and labelling. In order to overcome the limitations of image-based classification, a novel approach of augmenting neural network classification with positional data from lab based positional tracking systems was proposed. Ultrasound and positional data were collected from an abdominal phantom which allowed for the classification of six overlapping and hard to recognise abdominal cross sections with accuracies above 98%. A novel pilot study on 11 soft body Thiel cadavers, further refined this technique by exploring normalisation as a method to reduce the variability of coordinates produced when scanning the abdominal cavity and achieved an accuracy 96.8% using 3 points of normalisation. This work has demonstrated the efficacy of classification of abdominal ultrasound cross sections using neural networks and overcome the accuracy limitations of image-only classification of common ultrasound edge cases using a novel positional tracking approach, that achieved results far exceeding the current industry classification standards of abdominal cross sections.
Date of Award27 Mar 2024
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde & EPSRC (Engineering and Physical Sciences Research Council)
SupervisorGordon Dobie (Supervisor) & Rory Hampson (Supervisor)

Cite this