Author Of 1 Presentation
P0006 - Detecting treatment response on T1 gadolinium enhancing lesion burden in clinical trials of multiple sclerosis with deep learning (ID 998)
Abstract
Background
Ocrelizumab (OCR) is a humanized anti-CD20+ monoclonal antibody approved for the treatment of relapsing and primary progressive forms of multiple sclerosis (MS). It suppresses the development of new areas of inflammation as shown by the reduction in the number of T1 gadolinium (Gd) enhancing lesions. Deep learning (DL) based segmentation of lesions has the potential to automate these manual reads; enabling faster and more reproducible quantification.
Objectives
To evaluate a DL model’s ability to detect OCR treatment response on the number of T1 Gd enhancing lesions in clinical trials of relapsing remitting MS (RRMS).
Methods
We used images from Opera I trials (NCT01247324, n=898), to train a Unet model to segment both unenhancing and Gd-enhancing T1 lesions. T1 Gd- enhancing lesions in ground truth (GT) masks were present in ~44% of patients at baseline and ~15% across all time points. We created a dataset with an equal number of imaging volumes with and without enhancing lesions and used 70%-30% data splits for training and validation. The model was tested on images from Opera II trials (NCT01412333, n=905). To detect significant differences between treatment and control arms, we performed a negative binomial regression on the number of T1 Gd-enhancing lesions with baseline imaging and clinical covariates.
Results
The DL model achieved a mean dice coefficient (DC) of 0.72, lesion true positive rate of 0.92 (LTPR), lesion false positive rate (LFPR) of 0.06 and a volume correlation coefficient (CC) of 0.97 for Gd-enhancing lesion segmentation. For unenhancing lesion segmentation, the mean DC was 0.68, LTPR 0.78, LFPR 0.18 and volume CC of 0.97. The model had the highest false positives for lesions smaller than 10 voxels (voxel size: 1x1x3 mm3). For Gd-enhancing lesion segmentation a significant OCR treatment effect (p<0.001) in reducing the mean number of Gd-enhancing lesions at 24, 48 and 96 weeks (92%, 96%, 97% reduction from GT manual reads vs 67%, 71%, 78% from model predictions) was detected.
Conclusions
Our DL model performed Gd-enhancing lesion segmentation comparably to similar DL models in the literature and showed a high correlation to GT manual reads. Our model also had sufficiently high sensitivity to detect an OCR treatment response consistent with neuro-radiologist assessments. To our knowledge, this is the first study to report that a DL model has the sensitivity to detect treatment response on Gd-enhancing lesions.