The Effect of Image Stack Resampling on Manual Segmentation of the Putamen from Magnetic Resonance Imaging

Abstract

In neurology, brain region surface areas and volumes are often utilized to confirm diagnosis. These metrics are calculated from three-dimensional (3D), digital models sculpted from magnetic resonance imaging (MRI) stacks via a process called segmentation. Manual and automated segmentation highlights and in-fills 3D pixels (“voxels”), with length, height, and width dimensions dependent on the spatial increment of scan image capture. The capture of diagnostic images in clinical settings is necessarily low-resolution due to patient safety, time, and costs of equipment operation, yielding 3D surface areas and volumes contingent on a variable degree of missing anatomical information. The degree to which scan resolution affects 3D volumetric and diagnostic potential is poorly documented. We therefore investigate whether an image stack pre-processing technique known as “resampling” may improve volumetric precision in manual segmentation. Resampling is a common data-handling technique that increases the resolution of 3D digital images by inserting new planes of voxels with grayscale values interpolated from existing neighboring values. To form preliminary recommendations regarding the efficacy of resampling procedures for ameliorating low scan resolutions, we quantify inter- and intra-user variation in manual segmentation of the putamen from an MRI study of a healthy individual, using two protocols: 1) original scan resolution, and 2) up-sampled by 100% in all three dimensions, in a repeated-measures experimental design. Six participants (three per test group) segmented the left and right putamina five successive times using a standardized segmentation protocol. For each model surface area and volume data were collected. Results indicate that volume is significantly less variable for putamina that were segmented from the resampled dataset when compared to those segmented from the original dataset. The inverse result was found for surface area: variation was significantly lower for the original dataset. Intra- and inter-user replicability did not differ between methods. Overall, these results support the utilization of up-sampling to increase the accuracy of repeated volume measurements.