Comparing lesion segmentation methods in multiple sclerosis: Input from one manually delineated subject is sufficient for accurate lesion segmentation

Purpose: Accurate lesion segmentation is important for measurements of lesion load and atrophy in subjects with multiple sclerosis (MS). International MS lesion challenges show a preference of convolutional neural networks (CNN) strategies, such as nicMSlesions. However, since the software is trained on fairly homogenous training data, we aimed to test the performance of nicMSlesions in an independent dataset with manual and other automatic lesion segmentations to determine whether this method is suitable for larger, multi-center studies. Methods: Manual lesion segmentation was performed in fourteen subjects with MS on sagittal 3D FLAIR images from a 3T GE whole-body scanner with 8-channel head coil. We compared five different categories of automated lesion segmentation methods for their volumetric and spatial agreement with manual segmentation: (i) unsupervised, untrained (LesionTOADS); (ii) supervised, untrained (LST-LPA and nicMSlesions with default settings); (iii) supervised, untrained with threshold adjustment (LST-LPA optimized for current data); (iv) supervised, trained with leave-one-out cross-validation on fourteen subjects with MS (nicMSlesions and BIANCA); and (v) supervised, trained on a single subject with MS (nicMSlesions). Volumetric accuracy was determined by the intra-class correlation coefficient (ICC) and spatial accuracy by Dice's similarity index (SI). Volumes and SI were compared between methods using repeated measures ANOVA or Friedman tests with post-hoc pairwise comparison. Results: The best volumetric and spatial agreement with manual was obtained with the supervised and trained methods nicMSlesions and BIANCA (ICC absolute agreement > 0.968 and median SI > 0.643) and the worst with the unsupervised, untrained method LesionTOADS (ICC absolute agreement = 0.140 and median SI = 0.444). Agreement with manual in the single-subject network training of nicMSlesions was poor for input with low lesion volumes (i.e. two subjects with lesion volumes ≤ 3.0 ml). For the other twelve subjects, ICC varied from 0.593 to 0.973 and median SI varied from 0.535 to 0.606. In all cases, the single-subject trained nicMSlesions segmentations outperformed LesionTOADS, and in almost all cases it also outperformed LST-LPA. Conclusion: Input from only one subject to re-train the deep learning CNN nicMSlesions is sufficient for adequate lesion segmentation, with on average higher volumetric and spatial agreement with manual than obtained with the untrained methods LesionTOADS and LST-LPA.