Quality Control Methods for the Automatic Segmentation of White Matter Hyperintensities from Brain MRI using Deep Learning

Collaborating institution: Aicura Medical (Berlin, Germany)

About the project

White matter hyperintensities (WMH) are common in ageing and a main signature of sporadic small vessel disease and vascular dementia. Of perhaps similar aetiology, but undoubtedly with similar appearance, are brain lesions in multiple sclerosis. Not surprisingly, considerable efforts have been dedicated worldwide to assess them automatically, for achieving better diagnosis and interventions for patients with cardiovascular and multiple sclerosis diseases. Many automated methods have been proposed for assessing these neuroradiological features, most of them using convolutional neural networks (CNN), as they generally produce better results than conventional machine learning algorithms. In a clinical setting it is important to understand the constraints and instabilities of a CNN model and to assess the quality of the results being reported. Whilst manual quality control on a large scale is not attainable, automated methods have been developed for this purpose. This project explores the feasibility of applying the most promising automated quality control methods for CNN-based segmentation models to the task of WMH segmentation.