Author Of 1 Presentation
P0016 - Progressive multifocal leukoencephalopathy lesion and brain parenchymal segmentation from MRI using serial deep convolutional neural networks (ID 1531)
Abstract
Background
Progressive multifocal leukoencephalopathy (PML) is a rare opportunistic brain infection caused by the JC virus associated with significant morbidity and mortality, which can occur in the context of certain MS disease modifying therapies. There are currently no validated automatic methods for quantification of PML lesion burden or brain atrophy on MRI.
Objectives
We assessed whether deep learning techniques can be employed for automated brain parenchymal and lesion segmentation in PML using an approach dubbed “JCnet,” named after the causative viral agent.
Methods
We performed a retrospective analysis of PML patients who were evaluated at the NIH Neuroimmunology Clinic. MRI scans were acquired on either a Siemens Skyra or a Philips 3T MRI scanners. For PML brain and lesion segmentation, we implement a 3D patch-based approach with two consecutive fully convolutional neural networks (CNNs) with a feature pyramid architecture. The first network performs brain extraction as foreground, with meninges and cerebrospinal fluid spaces as background , while the second segments the underlying PML lesion(s). We measured the segmentation accuracy using Dice similarity coefficient (DSC) and absolute volume differences (AVD). We evaluated JCnet against methods designed for normal-appearing brain segmentation, FSL/FMRIB's Automated Segmentation Tool (FAST) and FreeSurfer, as well as MS lesion segmentation, Lesion Segmentation Toolbox (LST) and Lesion-TOpology-preserving Anatomical Segmentation (LTOADS). Comparisons were performed using Wilcoxon matched-pairs signed-ranks test.
Results
A total of 41 PML patients (mean age 55 years, SD 13; 44% female) were included in the analysis. The cohort was empirically divided into 31 training and 10 testing cases sampled at random. The mean time between PML onset and MRI acquisition was 4.5 months (range 0.6 – 44.5 months). JCnet resulted in a 4% and 64cm3 absolute improvement in DSC and AVD compared to FAST (p=0.005 and 0.01), and a 6% and 41cm3 absolute improvement compared to FreeSurfer respectively (p=0.005 and p=0.02). This was driven in part by improved segmentation of brain tissue within T1-hypointense PML lesions. For PML lesion segmentation, there was an absolute improvement of 42% and 14cm3 in DSC and AVD respectively compared to LST, and 53% and 19cm3 absolute improvements compared to LTOADS respectively (p=0.005 for all lesion comparisons). This was driven by improved sensitivity of supra- and infratentorial PML lesion identification and segmentation.
Conclusions
We employ an end-to-end deep learning-based method for automated segmentation of lesion and brain parenchymal volume in PML. By tracking quantitative measurements of PML-related MRI changes, this approach provides a window for clinicians and scientists to accurately monitor PML radiographically and its response to experimental therapies.