Author Identifier

Brandon Abela

https://orcid.org/0000-0001-5764-9554

Date of Award

2024

Document Type

Thesis

Publisher

Edith Cowan University

Degree Name

Master of Computing and Security by Research

School

School of Science

First Supervisor

Jumana Abu-Khalaf

Second Supervisor

Martin Masek

Abstract

Diabetic Foot Disease (DFD) costs the Australian healthcare system $1.6 billion annually. A type of DFD is Diabetic Foot Osteomyelitis (DFO) which is a bone infection. Diagnosis of DFO currently involves an initial X-ray of the foot examined by a clinician, but due to the poor accuracy of this technique (68.91% sensitivity; 77.99% specificity), in many cases, a second[1]line approach is required. Typically, suspicion needs to be confirmed with more time consuming and expensive methods, such as probe-to-bone test or MRI (95.76% sensitivity; 81.79% specificity), which slows down treatment acquisition and results in unnecessary costs. A potential solution to the low sensitivity of X-ray-based DFO diagnosis, is the integration of Deep Learning (DL) into the diagnostic process. However, developing DL models for clinical deployment requires the model to be resistant to feature distribution shifts that naturally occur when transferring a model to a new context. To ensure that a model is robust and viable for clinical deployment, the features present during development need to be identified. This allows for the isolation of specific features and the evaluation of their impact on model performance. The exploration of the features present in a DFO dataset makes up the first stage of this work. After identifying the features contributing to model performance, it is then necessary to control the features negatively impacting performance within the context of a distribution shift. Features that negatively affect performance in this context are called spurious features, which are features unrelated to the target class that co-occur with the target class in the training dataset. The application of data balancing via filtering to mitigate spurious correlations is a common approach. However, this limits the data that can be used, further exacerbating a problem in medical AI research, limited labelled data. Therefore, the second part of this study introduces and evaluates a data balancing process for validating the existence of spurious correlations stemming from imbalanced data. This process is shown to be effective at validating spurious correlations and is used to aid the development of the final DFO classification model. The developed model was evaluated both free from and under the conditions of a forced distribution shift and maintained similar performance, insinuating that the model had been effectively developed with a resistance to a defined distribution shift. Overall, this study made a range of contributions including the identification of co-morbid abnormalities in a DFO dataset, the development of a process for validating spurious features and development of an effective DFO classification model. The major findings of this thesis include the following: amputations and imaging procedures are identified and confirmed to be spurious features affecting model performance; pair matching can be used during testing to validate spurious features; and extracting a patch from the region of interest is an effective method for developing a DFO classification model. This work has the potential to reduce radiologists’ workload, improve case turnaround time, increase access to accurate diagnostic tools for low socio-economic regions, improve the healthcare systems return on investment, and improve reliable AI deployment by ensuring models behave as expected.

DOI

10.25958/97zk-jd56

Access Note

Access to this thesis is embargoed until 26th November 2025

Available for download on Wednesday, November 26, 2025

Share

 
COinS