Robust Identification of White Matter Hyperintensities in Uncontrolled Settings Using Deep Learning

Alice Schiavone, Sebastian Nørgaard Llambias, Jacob Johansen, Silvia Ingala, Akshay Pai, Mads Nielsen, Mostafa Mehdipour Ghazi

Keywords: Deep learning, domain shift, data augmentation, white matter hyperintensity

TL;DR: We propose a method for the segmentation of white matter hyperintensities in fluid-attenuated inversion recovery (FLAIR) scans, which is robust to domain shift compared to state-of-the-art methods.

Abstract: White matter hyperintensities (WMH) are associated with an increased risk of stroke, cognitive decline, and dementia. A robust, yet accurate detection of WMH can help with the prevention of more lesions from forming. The task is still challenging as the lesions are often small and irregular. Hence, we propose a robust deep learning-based method for the automatic segmentation of WMH only using fluid-attenuated inversion recovery (FLAIR) scans and MRI-specific data augmentation and compare it with state-of-the-art methods. The methods are tested on public and private data, and we show that our model is more robust to domain shift and achieves higher segmentation accuracy than the alternatives.