Automatic rat brain image segmentation using triple cascaded convolutional neural networks in a clinical PET/MR

Abstract The purpose of this work was to develop and evaluate a deep learning approach for automatic rat brain image segmentation magnetic resonance imaging (MRI) images in clinical PET/MR, providing useful tool analyzing studies the pathology progression neurological disease validate new radiotracers therapeutic agents. Rat PET/MR ( N = 56) were collected from system using dedicated small-animal phased array coil. A method based on triple cascaded convolutional neural network (CNN) developed, where, rectangular region interest covering whole brain, entire volume outlined CNN, then fed into segment both cerebellum cerebrum, finally sub-cortical structures within cerebrum including hippocampus, thalamus, striatum, lateral ventricles prefrontal cortex segmented out last CNN. dice score coefficient (DSC) between manually drawn labels predicted used quantitatively accuracy. proposed achieved mean DSC 0.965, 0.927, 0.858, 0.594, 0.847, 0.674 0.838 cerebellum, ventricles, respectively. Compared with results reported previous publications atlas-based methods, demonstrated improved performance segmentation. In conclusion, high accuracy MRI enabled possibility processing small animal research.