A lesion-level deep learning approach to predict enhancing lesions from nonenhanced images in multiple sclerosis

Background: In patients with multiple sclerosis (MS), contrast enhancing lesions are an important marker of disease activity. Identifying these on non-enhanced images would potentially allow omitting contrast agents in MR examinations of MS patients. Methods: In total, 485 consecutive patients with 1034 exams from a prospective observational cohort of multiple sclerosis patients were included in this analysis (scanned between 01/2008 and 09/2017). 10,311 lesion-level patches were generated from 297 patients to train a deep convolutional neural network (85% training, 15% validation) to discriminate between enhancing and non-enhancing lesions. Manual annotation of individual lesions enhancement on T1-weighted images after contrast administration served as the ground truth. This network was then tested on the remaining 188 patients. An external data set of 29 patients was additionally used to validate results. Network performance was statistically assessed by AUC, sensitivity, specificity and accuracy analysis. Results: MR examinations of 485 patients were analyzed. Out of this dataset, 297 patients (mean age: 34.8 years [+ or -] 9.7, ratio female:male 196:101) were used for training and validating the network and 188 patients (mean age: 33.9 years [+ or -] 9.1, ratio female:male 130:58) were used for testing. On the lesion level, an AUC of 0.905 was observed. A prespecified probability threshold of 0.75 allowed for identification of enhancing lesions with a sensitivity of 47.65%, specificity of 97.27% and accuracy of 95.27%. On patient level, the network exhibited a sensitivity of 82.76%, specificity of 66.15% and accuracy of 71.28%. Data from an external center confirmed these results. Conclusions: A deep learning approach at the lesion level is effective in predicting contrast-enhancing lesions from nonenhanced MRI images of patients with MS.