AI Adoption in Real-World Clinical Neuroimaging Applications: Practical Challenges and Solutions

Abstract

Deep learning has demonstrated a capacity to revolutionise human life in the last several years. Medical imaging, which has a vast data footprint, has emerged as a pioneering area that has seen rapid adoption of deep learning approaches to the domain. The development of new imaging-based algorithms is directed toward improved efficiency and accuracy of disease diagnosis and prognosis; and enhanced disease progression monitoring. These models also have the potential to provide insights into disease pathomechanisms.

However, the translation rate of newly described deep learning models into real-world clinical practice is extraordinarily low, despite an exponential increase in the number of research publications over the past few years. The cost of of data collection and annotation required to achieve model performance sufficient for clinical use; and to provide persuasive evidence of utility in real-world settings are significant roadblocks.

This thesis investigates solutions to the challenges of adapting deep neural networks to real-world settings. To improve the performance of algorithms, while reducing the costs of meticulous labelling, a novel model Masked Multi-Task Network is proposed for classification using only case-level labels; and a new training approach is proposed to tackle the issue of noisy labels in a federated learning setting. Furthermore, an in-depth analysis of the requirements for sample size used for training is conducted, to guide the development of deep learning models for large-scale adoption.

The research presented in this thesis encompasses the clinical validation and technical steps required for the commercialisation of two exemplary neuroimaging deep learning algorithms based on above works. This work also offers valuable insight into the compilation of requisite documentation for medical device registration, providing a valuable resource for researchers who wish to translate their models from the bench to the bedside.