Ex vivo MRI is increasingly popular in human brain atlasing. The absence of motion artifacts allows for long acquisitions that yield ultra-high resolution images [1]. Most of these studies rely on animal scanners and specialized coils that 1) are not widely available; and 2) cannot accommodate whole human brains, which is desirable to study human-specific disorders. Achieving ultra-high resolution with clinical scanners is feasible, but requires 3D acquisition sequences whose requirements quickly exceed the hardware capabilities of the scanner (e.g., RAM memory). Such limitations can be circumvented with multi-slab MRI, but imperfections in RF pulse profiles and flip angles across the slab thickness create slab boundary artifacts (SBA) (a.k.a. the Venetian Blind artifact), which bias subsequent analyses of the scans. Even though SBA can be mitigated during acquisition with slice oversampling [2], software post-processing is desirable in order to decrease the amount of required slab overlap and to remove effects that could not be corrected in the acquisition.

We propose a method for correction of SBA and bias field (BF); since both artifacts cause signal loss, simultaneous correction produces better results. A generative model of brain anatomy [3] is proposed based on the linear alignment of a voxelwise probabilistic atlas (prior) to the image. Given a voxelwise segmentation produced by the atlas, the observed log-transformed image intensities are assumed to be conditionally independent samples of different Gaussian mixture models, as indexed by the segmentation, and further corrupted by SBA and BF. These artifacts are modeled as a linear combination of basis functions that is added to the signal (the effect is thus multiplicative in the natural domain). Within this framework, estimating the SBA and BF is equivalent to finding the most likely linear coefficients, given the observed image and the probabilistic atlas. Using Bayes’ rule, this problem can be written:

$$ \arg\max_{\{c,\theta\}} p(i | c,\theta) p(c) p(\theta), $$

where $$$c$$$ are the SBA/BF coefficients and $$$\theta$$$ are the Gaussian parameters. This objective function is optimized with a Generalized Expectation Maximization algorithm [3]. An independent set of 2D, fourth order polynomials for each slice in the volume is used to model the SBA and BF artifacts. The number of components of the Gaussian mixtures was set to 2 for all tissue types.

We acquired MRI data from two postmortem cases on a 3T Siemens Trio with a multi-slab bSSFP sequence (TE/TR=5.3/10.6ms, flip angle $$$35^\circ$$$, 4 axial slabs with 112 slices each, 57% slice oversampling, 0.25 mm isotropic voxels). Four RF increments (0,90,180, 270 degrees) were averaged to reduce banding artifacts. The protocol was repeated 10 times to increase the SNR (total acquisition time: 60 hours). For the evaluation, the cerebral white matter was manually delineated in 10 equispaced coronal slices of each scan.

We compared our method with a combination of Kholmovski's SBA correction algorithm [4] and N4 BF correction [5] using two metrics: the coefficient of variation ($$$CV=\sigma/\mu$$$) of the white matter intensities, which measures the success of the BF correction, and the Hellinger distance between the distributions of white matter intensities at the slab boundaries and in the center of the slabs, which is a proxy for the quality of the SBA correction.

Figure 1 shows the mean $$$CV$$$ and $$$H$$$ for the different approaches. The baseline approach successfully reduces both metrics compared with the baseline, and the proposed method further improves both metrics. Even though the gain might seem modest at first in quantitative terms, the qualitative results in Figures 2-4 illustrate the superiority of our approach.

These results show that the proposed technique can produce high quality, ultra-high resolution MRI images using a clinical scanner. The method is efficient and runs in less than 5 minutes. Future work will include testing on a larger sample, combination with reconstruction-based SBA correction, and application to other modalities such as MR angiography and diffusion MRI.