Abstract Details

Title Achieving Human Level Performance for Automatic Lesion Segmentation in Multiple Sclerosis with Deep Learning

Topic Multiple Sclerosis

Presentation(s) P1 - Poster Session 1 (12:00 PM-1:00 PM)

Poster/Presentation
Number 9-011

Objective To develop deep learning based automatic lesion segmentation in multispectral brain MRI of multiple sclerosis (MS).

Background Accurate segmentation of white matter lesion in MS patients is key to quantitatively assessing disease status and to evaluate the effectiveness of treatments. Given the complexity and variability of lesion appearance in brain MRI, manual segmentation by radiologists is still regarded as ground truth; which, however, suffers from reader to reader variability, high cost, and low efficiency. To accelerate large-scale studies, fully automatic lesion identification and segmentation is desirable.

Design/Methods In this study we developed 2D U-Net deep learning models for automated T2 lesion segmentation in MS patients. We trained the models with multispectral brain MRI from a multi-site, multi-scanner, clinical trial dataset (NCT01247324) obtained from relapsing-remitting MS patients (N=898, 20% hold-out for validation). We tested the models on an independent dataset, which was obtained from another MS clinical trial (NCT01412333) that employed the same imaging protocol (N=905).

Results In comparison with radiologist segmentation, the multispectral U-net model, combining T2-weighted, T1-weighted, and FLAIR images, produces the best performance of the various combinations of input images evaluated. We have achieved mean DICE coefficient of 0.83 for lesion volumes larger than 15 ml, 0.78 for lesion volumes between 5 and 15 ml, and 0.67 for lesion volumes smaller than 5 ml. In addition, estimates of total lesion volume based on U-net segmentation are highly correlated with that of radiologist segmentation in the test dataset (R=0.96, p<0.00001).

Conclusions The DICE coefficients in our test are at the same level as those comparing segmentations between radiologists in the literature. Our results demonstrate that deep learning methods hold great promise to accelerate lesion-based analysis in large scale MS studies.

Authors/Disclosures
PRESENTER	No disclosure on file
David Clayton	David Clayton has received personal compensation for serving as an employee of Genentech/Roche. David Clayton has received stock or an ownership interest from Roche. An immediate family member of David Clayton has received publishing royalties from a publication relating to health care.
	No disclosure on file
	No disclosure on file
	No disclosure on file