Thesis abstract:

The rapidly expanding development of AI computer aided diagnosis (CAD) tools and their deployment into the healthcare space will disrupt current workflow practices of large number of physicians.

Many studies have analysed the issues which contribute to the successful deployment of AI techniques in healthcare. Success development and deployment requires navigation of a pathway with many components.

The aim of this thesis is to specifically examine issues relating to the decision-making processes physicians will need to undertake in the altered workplace environments where CAD tools have been deployed. The thesis will add to work done previously exploring the ‘acceptability’ of AI CAD in healthcare.

The thesis will specifically utilise outputs from a CAD tool currently deployed in a large radiology provider for TB detection and the related patient data from my Supervisor B, Dr Jin-Gun Cho. I plan to develop previously described statistical techniques to explore the effect of several input factors on the accuracy of TB detection. Specifically, I will be examining:

• the pattern of presentation of CAD data in the existing workflow
• the role of weighted risk of False negatives by the physicians
• the role of referred 2nd opinion specialists and their place in the workflow and the
• cutoff values in the ROCs used by the CAD tool.

It is proposed that radiologist be enrolled in a series of experiments to examine the above variables.

The outputs will be examined using techniques of applied Bayesian statistics such as Decision Curve Analysis and Application Area Analysis to examine the variation in specificity and sensitivity related to interactions between these inputs. In addition, these results may guide discussion on the possible hazards and benefits of specific input factors.