Quantification of errors in volume measurements of the caudate nucleus using magnetic resonance imaging.

PURPOSE: To quantify the various sources of error in measuring the volume of the caudate nucleus and to understand these errors would lead to the standardization of the MRI protocol and would make the utility of data from around the world more viable in a global database.
MATERIALS AND METHODS: We collected data at four different sites all using a Siemens 1.5T Vision MR Scanner. In all cases the same 3D gradient-echo scans were used on a single volunteer and analyzed by a set of five observers.
RESULTS: The errors estimated were: system calibration (a random variation of up to 1.2%), partial volume error (a bias of up to 1.5% using isotropic resolution of 1 x 1 x 1 mm(3)), geometric distortion (a potential bias of 1%), intra-observer error (a random variation of up to 3%), effects of ringing (area biases of up to 7% when a zoom of 4 was used) and inter-observer error (with a bias of usually 5- 10% but sometimes as large as 16% among our five observers). Individual mean variations from one system to another differed by less than 5% (except for two observers at one site), consistent with a maximum error of 7% coming from the area bias due to limitations in the images themselves. We also measured the effect of variable resolution on the volume estimates and found that the measured volumes were consistent over a broad range of signal-to-noise-ratios (SNRs).
CONCLUSION: Given the observed dependence of the caudate volumes on SNR and resolution, if isotropic resolution is required because a complicated structure is being imaged, then the lower SNR suffered by collecting 1 x 1 x 1 mm(3) data at 1.5T still appears to be sufficient to make accurate volume measurements as long as the contrast-to-noise ratio (CNR) is on the order of 4:1. Based on our results, predictions are made as to what the best approach would be to improve the data acquisition scheme to keep individual errors under 2% and biases under 3.5%. We conclude that if users can be trained to identify the structure of interest in the same way, the inter-observer error could be reduced to that of intra-observer day-to-day error. Copyright 2002 Wiley-Liss, Inc.