Automatic Rodent Brain MRI Lesion Segmentation with Fully Convolutional Networks

Manual segmentation of rodent brain lesions from magneticresonance images (MRIs) is an arduous, time-consuming and subjectivetask that is highly important in pre-clinical research. Several automaticmethods have been developed for different human brain MRI segmen-tation, but little research has targeted automatic rodent lesion segmen-tation. The existing tools for performing automatic lesion segmentationin rodents are constrained by strict assumptions about the data. Deeplearning has been successfully used for medical image segmentation. How-ever, there has not been any deep learning approach specifically designedfor tackling rodent brain lesion segmentation. In this work, we proposea novel Fully Convolutional Network (FCN), RatLesNet, for the afore-mentioned task. Our dataset consists of 131 T2-weighted rat brain scansfrom 4 different studies in which ischemic stroke was induced by transientmiddle cerebral artery occlusion. We compare our method with two other3D FCNs originally developed for anatomical segmentation (VoxResNetand 3D-U-Net) with 5-fold cross-validation on a single study and a gener-alization test, where the training was done on a single study and testingon three remaining studies. The labels generated by our method werequantitatively and qualitatively better than the predictions of the com-pared methods. The average Dice coefficient achieved in the 5-fold cross-validation experiment with the proposed approach was 0.88, between3.7% and 38% higher than the compared architectures. The presentedarchitecture also outperformed the other FCNs at generalizing on differ-ent studies, achieving the average Dice coefficient of 0.79.