Automatic rat brain image segmentation using triple cascaded convolutional neural networks in a clinical PET/MR
Ya Gao,Zaisheng Li,Cheng Song,Lei Li,Mengmeng Li,Jeffrey Schmall,Hui Liu,Jianmin Yuan,Zhe Wang,Tianyi Zeng,Lingzhi Hu,Qun Chen,Yanjun Zhang
DOI: https://doi.org/10.1088/1361-6560/abd2c5
IF: 3.5
2021-02-01
Physics in Medicine and Biology
Abstract:Abstract The purpose of this work was to develop and evaluate a deep learning approach for automatic rat brain image segmentation of magnetic resonance imaging (MRI) images in a clinical PET/MR, providing a useful tool for analyzing studies of the pathology and progression of neurological disease and to validate new radiotracers and therapeutic agents. Rat brain PET/MR images ( N = 56) were collected from a clinical PET/MR system using a dedicated small-animal imaging phased array coil. A segmentation method based on a triple cascaded convolutional neural network (CNN) was developed, where, for a rectangular region of interest covering the whole brain, the entire brain volume was outlined using a CNN, then the outlined brain was fed into the cascaded network to segment both the cerebellum and cerebrum, and finally the sub-cortical structures within the cerebrum including hippocampus, thalamus, striatum, lateral ventricles and prefrontal cortex were segmented out using the last cascaded CNN. The dice score coefficient (DSC) between manually drawn labels and predicted labels were used to quantitatively evaluate the segmentation accuracy. The proposed method achieved a mean DSC of 0.965, 0.927, 0.858, 0.594, 0.847, 0.674 and 0.838 for whole brain, cerebellum, hippocampus, lateral ventricles, striatum, prefrontal cortex and thalamus, respectively. Compared with the segmentation results reported in previous publications using atlas-based methods, the proposed method demonstrated improved performance in the whole brain and cerebellum segmentation. In conclusion, the proposed method achieved high accuracy for rat brain segmentation in MRI images from a clinical PET/MR and enabled the possibility of automatic rat brain image processing for small animal neurological research.
engineering, biomedical,radiology, nuclear medicine & medical imaging