Abstract:
Multiple sclerosis (MS) affects about 3 million people worldwide, and is typically diagnosed in young adults. The disease affects the central nervous system, where it causes inflammation of nerve tissue and gradual degeneration of the nerve cells. The main target of MS is the insulation layer surrounding the nerve cells, reducing or even preventing signal transmission. The damage is visible on magnetic resonance (MR) images of the brain and spinal cord as regions of inflammation (lesions) and loss of brain tissue (atrophy). MS leads to a wide range of symptoms including cognitive impairment, which is present in up to 70% of people with MS.
Radiological findings however do not always manifest as clinical symptoms and the other way around, known as the “clinico-radiological paradox”. The main goal of this thesis was to offer new insights in this paradox for cognitive problems using artificial intelligence (AI). However, databases with brain images and cognitive information are scarce, impeding AI research. This thesis proposes three solutions for this problem: (1) facilitating data collection with digital cognitive tests, (2) reducing the need for large databases by using models that are trained to perform a related task (transfer learning) and (3) increasing accessibility to clinical datasets for AI modelling using federated learning.
First, icognition was presented. This is a smartphone-based application with three cognitive tests, designed to screen for impairment in the two most commonly affected cognitive domains in MS: memory and information processing speed. The application was shown to be reliable and valid, although the results should be confirmed in a cognitively impaired MS sample. The application allows screening for cognitive problems at home, thereby picking up cognitive deterioration early on. Furthermore, digitalising cognitive tests facilitates the creation of large research databases in the future. Second, “brain age”, interpretable as “how old the brain looks”, was explored as an in-between step to predict cognitive functioning. A model was trained to predict age, i.e. brain age, from brain MR images. The model overestimated age in people with MS, confirming that people with MS have older looking brains. Brain age moreover correlated with their information processing speed. Subsequently, a deep learning brain age model, trained on a large database of healthy people, was fine-tuned to predict cognition on a smaller MS database (transfer learning). Although the final model performed rather poorly at predicting cognitive performance from brain MRI, we proved in the same study that the model could be trained without sharing data between clinical centres. Instead of first collecting all data at one place, the model was sent to the data, where it was updated locally (federated learning). In this way, the model exhibited learning behaviour at each clinical centre, setting the stage for training better models without sharing sensitive clinical data such as brain MRI.
This thesis explored solutions for AI research in a context of low data availability. Reaching large and high-quality data sets could eventually enable AI to help patients with MS and their caregivers managing an unpredictable and burdensome disease.