Convolutional neural networks for the segmentation of small rodent brain MRI

Image segmentation is a common step in the analysis of preclinical brain MRI, often performed manually. This is a time-consuming procedure subject to inter- and intra- rater variability. A possible alternative is the use of automated, registration-based segmentation, which suffers from a bias owed to the limited capacity of registration to adapt to pathological conditions such as Traumatic Brain Injury (TBI). In this work a novel method is developed for the segmentation of small rodent brain MRI based on Convolutional Neural Networks (CNNs). The experiments here presented show how CNNs provide a fast, robust and accurate alternative to both manual and registration-based methods. This is demonstrated by accurately segmenting three large datasets of MRI scans of healthy and Huntington disease model mice, as well as TBI rats. MU-Net and MU-Net-R, the CCNs here presented, achieve human-level accuracy while eliminating intra-rater variability, alleviating the biases of registration-based segmentation, and with an inference time of less than one second per scan. Using these segmentation masks I designed a geometric construction to extract 39 parameters describing the position and orientation of the hippocampus, and later used them to classify epileptic vs. non-epileptic rats with a balanced accuracy of 0.80, five months after TBI. This clinically transferable geometric approach detects subjects at high-risk of post-traumatic epilepsy, paving the way towards subject stratification for antiepileptogenesis studies.