The bootstrap and cross-validation in neuroimaging applications: estimation of the distribution of extrema of random fields for single volume tests, with an application to ADC maps.

We discuss the assessment of signal change in single magnetic resonance images (MRI) based on quantifying significant departure from a reference distribution estimated from a large sample of normal subjects. The parametric approach is to build a test based on the expected distribution of extrema in random fields. However, in conditions where the variance is not uniform across the volume and the smoothness of the images is moderate to low, this test may be rather conservative. Furthermore, parametric tests are limited to datasets for which distributional assumptions hold. This paper investigates resampling methods that improve statistical tests for signal changes in single images in such adverse conditions, and that can be used for the assessment of images taken for clinical purposes. Two methods, the bootstrap and cross-validation, are compared. It is shown that the bootstrap may fail to provide a good estimate of the distribution of extrema of parametric maps. In contrast, calibration of the significance threshold by means of cross-validation (or related sampling without replacement techniques) address three issues at once: improved power, better voxel-by-voxel estimate of variance by local pooling, and adaptation to departures from ideal distributional assumptions on the signal. We apply the cross-validated tests to apparent diffusion coefficient maps, a type of MRI capable of detecting changes in the microstructural organization of brain parenchyma. We show that deviations from parametric assumptions are strong enough to cast doubt on the correctness of parametric tests for these images. As case studies, we present parametric maps of lesions in patients suffering from stroke and glioblastoma at different stages of evolution.