Evaluating the Performance of an Atlas-Free Automated Method of Brain Classification for Lesion Segmentation in Progressive Multifocal Leukoencephalopathy
Hadar Kolb1, Daniel Reich1, Irene Cortese2, Govind Nair2
1Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke, 2Neuroimmunology Clinic, National Institute of Neurological Disorders and Stroke, National Institutes of Health
Objective:
To explore the feasibility of an automated approach for segmentation and quantification of white matter lesions in the context of Progressive Multifocal Leukoencephalopathy (PML).
Background:
PML is a rare and often fatal infection caused by JC polyomavirus that almost exclusively occurs in people with impaired cellular immunity.  Magnetic resonance imaging (MRI) demonstrates typical multifocal hyperintense lesions on T2-weighted images affecting subcortical white matter and posterior fossa structures. MRI is helpful for monitoring PML disease course and development of an imaging-based quantitative measure of disease burden as an outcome measure, would be of great value. Therefore, we sought to develop an automated method of brain classification using derivative-based features (C-DEF) optimized for segmentation of PML lesions.
Design/Methods:
Using serial MRIs from a PML test subject, white matter lesions were manually segmented on T1-weighted and fluid-attenuated inversion recovery (FLAIR) images. CDEF was trained using these ‘gold-standard’ brain segmentation masks to identify healthy white and gray matter and PML lesions. Once CDEF was optimized on the test subject, the algorithm was applied to images obtained from 7 additional PML patients.
Results:
CDEF performed well on the test subject with average Dice coefficients between CDEF and gold-standard classification of 0.73 for lesions, 0.97 for grey matter, and 0.99 for white matter. Application to the larger cohort, however, was relatively poor. Generally, supratentorial lesions were accurately segmented, including small foci. However, CDEF often misclassified periventricular regions, cerebellar regions and healthy gray matter as lesions.
Conclusions:
Rapid and reliable quantification of PML lesion burden is desirable for monitoring disease progression and stabilization of PML, yet, this remains challenging due to the non-uniform, widespread nature of PML lesionsPreliminary results demonstrate that CDEF could be developed into a useful tool to measure disease burden by combining different MRI sequences to improve the segmentation quality of CDEF.