This study proposes a deep learning-based model that predicts white matter hyperintensity (WMH) using only three-dimensional(3D) carotid Time-of-Flight (TOF) magnetic resonance angiography (MRA) images as input. WMH is a representative imaging marker ...
This study proposes a deep learning-based model that predicts white matter hyperintensity (WMH) using only three-dimensional(3D) carotid Time-of-Flight (TOF) magnetic resonance angiography (MRA) images as input. WMH is a representative imaging marker of cerebral small vessel disease(cSVD) and is known to be associated with cognitive impairment, gait disturbances, and increased risk of stroke and dementia. While previous studies have primarily focused on the statistical correlations between carotid morphological features and WMH, direct prediction of WMH from carotid MRA imaging has not been explored. To address this gap, this study adopts an end-to-end deep learning approach that directly utilizes raw TOF MRA images for WMH classification.
This study investigates two classification tasks: binary classification to detect the presence of WMH, and three-class classification to assess its severity. Several convolutional neural network (CNN)-based models—SFCN, ResNet10, and MedicalNet—and a Transformer-based model, MST, were trained and evaluated. Among these, the SFCN model demonstrated the best performance, achieving 81.8% accuracy and an AUC of 0.882 in binary classification, and 66.4% accuracy with an AUC of 0.847 in WMH severity classification. To interpret the model’s predictions, saliency maps and occlusion sensitivity analyses were performed. The saliency visualization revealed that the trained models commonly focused on the anatomical structures of the carotid artery, particularly around the carotid bifurcation. Occlusion sensitivity analysis further confirmed that this region played a crucial role in the prediction process. These findings suggest that vascular information contained in carotid MRA images can be meaningfully utilized for predicting WMH.
This study demonstrates the feasibility of predicting and classifying WMH using carotid TOF MRA as the sole imaging modality within a deep learning framework. Furthermore, it highlights the potential of carotid imaging as a non-invasive tool for the early assessment of cerebrovascular abnormalities, establishing a practical basis for WMH prediction without relying on brain MRI.