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

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• White matter hyperintensities (WMHs) are common in elderly individuals' MRIs and crucial for diagnosis and prognosis.
• Manual segmentation of WMHs is time-consuming and prone to inconsistencies.
• Existing automated methods, often supervised CNNs, require extensive manual data for training.

Purpose of the Study:

• To develop a novel, unsupervised Convolutional Neural Network (CNN) for segmenting white matter hyperintensities (WMHs).
• To enable accurate and efficient WMH quantification without the need for manual delineations.
• To provide a robust biomarker for clinical research and patient assessment.

Main Methods:

• A novel unsupervised CNN approach was developed for WMH segmentation.
• The method reconstructs MRI sequences using weighted sums of WMH and tissue segmentations.
• The trained network segments new images rapidly and robustly.

Main Results:

• The unsupervised CNN achieved accurate WMH segmentation.
• The method demonstrated effectiveness across diverse datasets and scanners.
• Performance was comparable to state-of-the-art supervised methods, without manual training data.

Conclusions:

• The proposed unsupervised CNN method offers a fast, accurate, and robust solution for WMH segmentation.
• This approach eliminates the need for manual delineations, overcoming limitations of supervised methods.
• It holds significant potential for clinical research and diagnostic applications in neuroimaging.