RatLesNetv2: A Fully Convolutional Network for Rodent Brain Lesion Segmentation

Purpose: Introduce and validate a fully convolutional neural network (ConvNet), named RatLesNetv2, for segmenting rodent brain lesions in MR images. RatLesNetv2 is trained end to end on three-dimensional images and, unlike other commonly-used approaches, it requires no pre-processing. Methods: RatLesNetv2 was evaluated and compared with three other ConvNets on an exceptionally large dataset comprising 916 T2-weighted rat brain MRI scans at nine different lesion stages. RatLesNetv2 architecture resembles an auto-encoder and it incorporates residual blocks that facilitate its optimization. Results: RatLesNetv2 obtained similar to higher Dice coefficient values than the other ConvNets and it produced much more realistic and compact segmentations with notably less holes and lower Hausdorff distance. Such segmentations also exceeded inter-rater agreement. Conclusion: RatLesNetv2 outperformed other ConvNets and compared favorably to human-made annotations. As a result, RatLesNetv2 could be used for automated lesion segmentation, reducing human workload and improving reproducibility of the segmentations. RatLesNetv2 is publicly available at https://github.com/jmlipman/RatLesNetv2.