(I123)FP-CIT imaging is used for differential diagnosis of clinically uncertain Parkinsonian Syndromes. Conventional reporting relies on visual interpretation of images and analysis of semi-quantification results. However, this form of reporting is associated with variable diagnostic accuracy results. The first half of this thesis clarifies whether machine learning classification algorithms, used as computer aided diagnosis (CADx) tool, can offer improved performance.
Candidate machine learning classification algorithms were developed and compared to a range of semi-quantitative methods, which showed the superiority of machine learning tools in terms of binary classification performance. The best of the machine learning algorithms, based on 5 principal components and a linear Support Vector Machine classifier, was then integrated into clinical software for a reporting exercise (pilot and main study).
Results demonstrated that the CADx software had a consistently high standalone accuracy. In general, CADx caused reporters to give more consistent decisions and resulted in improved diagnostic accuracy when viewing images with unfamiliar appearances.
However, although these results were undoubtedly impressive, it was also clear that a number of additional, significant hurdles remained, that needed to be overcome before widespread clinical adoption could be achieved.
Consequently, the second half of this thesis focuses on addressing one particular aspect of the remaining translation gap for (I123)FP-CIT classification software, namely heterogeneity of the clinical environment. Introduction of new technology, such as machine learning, may require new metrics, which in this work were informed through novel methods (such as the use of innovative phantoms) and strategies, enabling sensitivity testing to be developed, applied and evaluated.
The pathway to acceptance of novel and progressive technology in the clinic is a tortuous one, and this thesis emphasises the importance of many factors in addition to the core technology that need to be addressed if such tools are ever to achieve clinical adoption.
