Aarhus University Seal / Aarhus Universitets segl

Torben Ellegaard Lund

Thalamus segmentation from MP2RAGE: a comparative study

Research output: Contribution to conferenceConference abstract for conferenceResearchpeer-review

Introduction: MPRAGE has become one of the most widely used MRI sequences to obtain T1-weighted anatomical images. However, at high static field strengths the increased inhomogeneity of B1 fields leads to high intensity variations across the image (bias field). Recently, it was suggested to use two MPRAGE images with different inversion times to construct an MP2RAGE image, free of B1 bias problems [1]. Since this so-called MP2RAGE sequence is independent of B1 as well as M0 and T2*, the T1 image contrast is improved but also different from conventional MPRAGE images. Thus, existing segmentation methods may not work well with this new sequence. In this study we tested three different automatic methods for the important task of segmenting the thalamus from human brain MP2RAGE images. Methods: Twelve healthy control subjects (age range 19 – 38 years, two females) were scanned with a whole brain MP2RAGE sequence on a Siemens Magnetom Skyra 3T MRI system with a 32 channel head coil. Parameters were TR=5 s, TI1=0.7 s, TI2=2.5 s, α1=4°, α2=5° and acquired at a nominal, isotropic, resolution of 1mm (acquisition matrix: 240x256, 176 sagittal slices). An experienced neuro-radiologist manually traced the thalami of all subjects using in-house software. Images were separately processed with Freesurfer (version 5.3) [2] and a recent processing pipeline [3], where anatomical structures are segmented with ANIMAL [4]. Finally, a non-local means patch based segmentation method (SNIPE) was applied using the manual segmentations as library in a leave-one-out fashion [5, 6]. All parameters were set to default except the search window for SNIPE, which was reduced to 7x7x7 voxels. This decreased the processing time without impairing the segmentation accuracy. Automatic thalamus segmentations from the three applied methods were compared to the manual segmentations using Dice similarity index (DSI), false-positive rate (FPR), and false-negative rate (FNR). Results: Three subjects failed the Freesurfer pipeline and were thus excluded from all comparisons. Figure 1 shows manual and automatic segmentations of the same subject along with the difference map showing SNIPE segmentation errors (lower right image). Mean DSI of the 9 included subjects for the three evaluated methods were: Freesurfer = 0.77±0.06, ANIMAL = 0.75±0.02, and SNIPE = 0.89±0.01. DSI were significantly (p<0.0001, two-tailed t-test) higher for SNIPE compared to Freesurfer and compared to ANIMAL, while no difference was found between Freesurfer and ANIMAL (p=0.26) (Figure 2). In general, all methods tend to under-segment the thalamus with FNRs around 10-15%, while very little over-segmentation is found in the results (Figure 3). As expected, increasing the number of training images in the library of SNIPE improves segmentation accuracy (Figure 4). Even though the DSI seems to plateau around a library size of 10-11 images, increasing the library may improve the accuracy even further. Conclusions: Widely used atlas based segmentation methods, such as Freesurfer and ANIMAL, do not work well with the new MP2RAGE sequence without modifications. Non-local patch based segmentation methods are better suited for the task, which is demonstrated by the improved accuracy of SNIPE compared to Freesurfer and ANIMAL. In addition, 25% of test images failed the Freesurfer pipeline, which is a high failure rate compared to studies running Freesurfer on conventional MPRAGE images [7]. All evaluated methods under-segments the thalamus, which may be desirable from a clinical point of view, if segmentation masks are used for characterizing the structure using e.g. diffusion or perfusion parameters obtained from other MRI sequences. For volumetric studies parameters of SNIPE can be adjusted to balance the over- and under-segmentation ratios.
Original languageEnglish
Publication year11 Jun 2014
Number of pages1
Publication statusPublished - 11 Jun 2014
Event20th Annual Meeting of the Organization for Human Brain Mapping - Hamburg, Germany
Duration: 8 Jun 201414 Jun 2014
Conference number: 20


Conference20th Annual Meeting of the Organization for Human Brain Mapping

    Research areas

  • MRI|, Segmentation, Thalamus

See relations at Aarhus University Citationformats

ID: 81760926