Comparison of tissue segmentation performance between 2D U-Net and 3D U-Net on brain MR Images
Boyeong Woo, Myungeun Lee
Abstract
In this paper, we compare the tissue segmentation performance of the 2D U-Net and 3D U-Net in brain MR images including Alzheimer's disease. The Open Access Series of Imaging Studies dataset used in this experiment consists of a cross-sectional collection of 416 subjects aged 18 to 96. The final segmentations were classified into 4 classes: background, cerebrospinal fluid, gray matter, and white matter. In the experiment, 3D U-Net showed highly stable segmentation results. Particularly, the evaluation of the 2D/3D U-Net segmentation has shown that the 3D U-Net provided higher dice similarity coefficients (94.4±4.5% ~ 95.6±3.0%) and lower Hausdorff distances ( 7.5±3.2mm ~ 12.3±3.9mm) than the 2D U-Net ( 89.6±4.9% ~ 96.22±3.1% and 9.1±5.5mm ~ 14.4±7.4mm). And the 3D U-Net provided a better performance despite having less training examples to learn from. The simple image processing pipeline was an added advantage. However, the number of parameters and the amount of memory required was a major limitation of using 3D U-Net.