A Robust and Accurate Deep-learning-based Method for the Segmentation of Subcortical Brain: Cross-dataset Evaluation of Generalization Performance

Abstract

PURPOSE: To analyze subcortical brain volume more reliably, we propose a deep learning segmentation method of subcortical brain based on magnetic resonance imaging (MRI) having high generalization performance, accuracy, and robustness. METHODS: First, local images of three-dimensional (3D) bounding boxes were extracted for seven subcortical structures (thalamus, putamen, caudate, pallidum, hippocampus, amygdala, and accumbens) from a whole brain MR image as inputs to the neural network. Second, dilated convolution layers, which input information of variable scope, were introduced to the blocks that make up the neural network. These blocks were connected in parallel to simultaneously process global and local information obtained by the dilated convolution layers. To evaluate generalization performance, different datasets were used for training and testing sessions (cross-dataset evaluation) because subcortical brain segmentation in clinical analysis is assumed to be applied to unknown datasets. RESULTS: The proposed method showed better generalization performance that can obtain stable accuracy for all structures, whereas the state-of-the-art deep learning method obtained extremely low accuracy for some structures. The proposed method performed segmentation for all samples without failing with significantly higher accuracy (P < 0.005) than conventional methods such as 3D U-Net, FreeSurfer, and Functional Magnetic Resonance Imaging of the Brain's (FMRIB's) Integrated Registration and Segmentation Tool in the FMRIB Software Library (FSL-FIRST). Moreover, when applying this proposed method to larger datasets, segmentation was robustly performed for all samples without producing segmentation results on the areas that were apparently different from anatomically relevant areas. On the other hand, FSL-FIRST produced segmentation results on the area that were apparently and largely different from the anatomically relevant area for about one-third to one-fourth of the datasets. CONCLUSION: The cross-dataset evaluation showed that the proposed method is superior to existing methods in terms of generalization performance, accuracy, and robustness.