SegAE: Unsupervised white matter lesion segmentation from brain MRIs using a CNN autoencoder

Abstract

Highlights

• A convolutional neural network designed for unsupervised segmentation of tissue and white matter hyperintensities (WMHs) in brain MRIs.
• Evaluations were conducted on two distinct datasets comprising data from six different scanners, all with ground truth manual WMH label masks.
• The method compares favorably to existing WHM segmentation methods and is fast and robust to highly variable WMH lesion load and atrophy in the brains of elderly subjects.