Abstract
PURPOSE: To develop and validate deep learning-based automatic brain segmentation for East Asians with comparison to data for healthy controls from Freesurfer based on a ground truth. METHODS: A total of 30 healthy participants were enrolled and underwent T1-weighted magnetic resonance imaging (MRI) using a 3-tesla MRI system. Our Neuro I software was developed based on a three-dimensional convolutional neural networks (CNNs)-based, deep-learning algorithm, which was trained using data for 776 healthy Koreans with normal cognition. Dice coefficient (D) was calculated for each brain segment and compared with control data by paired t-test. The inter-method reliability was assessed by intraclass correlation coefficient (ICC) and effect size. Pearson correlation analysis was applied to assess the relationship between D values for each method and participant ages. RESULTS: The D values obtained from Freesurfer (ver6.0) were significantly lower than those from Neuro I. The histogram of the Freesurfer results showed remarkable differences in the distribution of D values from Neuro I. Overall, D values obtained by Freesurfer and Neuro I showed positive correlations, but the slopes and intercepts were significantly different. It was showed the largest effect sizes ranged 1.07-3.22, and ICC also showed significantly poor to moderate correlations between the two methods (0.498 ≤ ICC ≤ 0.688). For Neuro I, D values resulted in reduced residuals when fitting data to a line of best fit, and indicated consistent values corresponding to each age, even in young and older adults. CONCLUSION: Freesurfer and Neuro I were not equivalent when compared to a ground truth, where Neuro I exhibited higher performance. We suggest that Neuro I is a useful alternative for the assessment of the brain volume.