Machine learning-based detection of vascular pathologies in MRI images: A multi-task learning approach for identifying white matter hyperintensities and vascular cognitive impairment in FLAIR MRI images

Abstract

This thesis investigates the potential of using FLAIR imaging data and machine learning (ML) to identify patients with vascular cognitive disease (VCD). VCD is currently underdiagnosed, and diagnosis usually happens at the late stages of the disease. Currently, there are no treatments available for VCD, but if caught early, it is possible to stop the progression of the disease by, for example, lifestyle changes. Therefore, early detection of VCD is essential.

A total of 750 magnetic resonance imaging (MRI) scans from a cohort of 506 women and men between 50 and 79 years of age from the Gothenburg mild cognitive impairment (MCI) study were analyzed. The cohort included controls, preclinical participants (subjective cognitive impairment and MCI), and patients with Alzheimer’s disease (AD), subcortical small-vessel disease (SSVD), and mixed AD/SSVD. FLAIR sequences from a 1.5 Tesla MRI scanner were used, together with previously analyzed white matter hyperintensity (WMH) volumes acquired from the FreeSurfer 5.3 software.

The input to the model consisted of FLAIR MRI volumes, while the ground truth data included clinical diagnoses and WMH volume measurements. A multi-task learning (MTL) model based on ResNet18 was implemented, combining regression of WMH volumes and classification of patient diagnoses. The explainability method Guided Grad-CAM was implemented to visualize the regions of the MRI images that contributed most to the model’s predictions or classifications. The model was analyzed in a binary and two multi-classification cases. The regression head produced WMH volume estimates comparable to those obtained from FreeSurfer. The classification head achieved an F1-score of 0.5806 for binary classification, but did not yield satisfactory results using three-class classification and five-class classification, with F1-scores of 0.5381 and 0.4465, respectively. This thesis demonstrates that it is possible to quantify WMH volumes and identify patients with vascular cognitive impairment based on FLAIR images using ML.