Volume 13 Supplement 1

Abstracts of the 2011 SCMR/Euro CMR Joint Scientific Sessions

Open Access

A fully-automated statistical method for characterization of flow artifact presence in cardiac MRI

  • Sotirios A Tsaftaris1,
  • Xiangzhi Zhou2 and
  • Rohan Dharmakumar2
Journal of Cardiovascular Magnetic Resonance201113(Suppl 1):P45


Published: 2 February 2011


Flow artifacts in MR images can appear as ghosts within and outside the body cavity. Current approaches for optimizing sequences for suppressing such artifacts rely on expert scoring or on semi-automated methods for evaluation.


To test a fully-automated statistical image-processing method that can quantify the presence of flow artifacts. The method was evaluated against expert scoring in the setting of cine blood-oxygen-level-dependent (BOLD) MRI with different SSFP imaging strategies.


Imaging studies

Six healthy dogs were studied in a Siemens 1.5T scanner using three breath-held 2D SSFP cine sequences (Table 1) in the basal position where flow artifacts are most pronounced. Shared scan parameters: spatial_resolution=1.2x1.2x5mm3, flip_angle=70°, temporal_resolution=10-12ms, no parallel imaging.
Table 1

2D cine SSFP Imaging sequences and parameters used. Shared parameters in text.


BOLD sensitivity

Flow artifacts



Flow compensation



















Image processing

Each cine stack I(t) (t denotes trigger time) was loaded in Matlab and the per-pixel mean (IA) and variance (IV) were found across t. Initialized with a rectangle the size of IA, a contour was evolved, using a level-set approach, until it converged to the body-air interface, providing a binary mask M (M=1 for air). Image IR=IV/(IA)2 (pixel-wise division) was calculated and all values of IR in air (M=1) were collected to estimate the excess kurtosis (γ) of their distribution H. The metric QK=γ was used to quantify the presence of flow artifacts.

Data analysis

Three expert reviewers, blinded to sequence type, scored 16 studies for the presence of ghosts [1(least) to 5(most)]. ANOVA was used to test for differences in scores/metric among sequences. QK was correlated with the reviewers’ median choice (QH) to assess agreement.


Fig. 1 shows a representative case from a study acquired with sequence B. Fig. 2 shows bar plots of QK values and scores (QH) for each imaging sequence. Correlation coefficient between QK and QH was 0.7 (R2=0.49;P<0.01).
Figure 1

Images from a canine study acquired with sequence B, showing the presence of artifacts and steps of the proposed method. Left: a systolic image (22nd frame in 55 total frames) of the study, showing optimal BOLD sensitivity and significant flow artifacts. Middle: The IR image as defined in the methods indicating that artifact information is retained (only pixels in air are shown as found by the level-set segmentation): the arrows point to artifacts. Right: The distribution (H) of the IR values in air: bracket indicates IR values from ghost artifact regions (see arrows in middle image) that act as outliers. Due to the presence of these outliers the excess kurtosis for this stack was y=QK=21.

Figure 2

Bar plots (mean ± standard error) for QK (derived using the kurtosis-based method) and QH (the experts median choice per study) grouped by sequence type. Intervals on top indicate statistical significance of individual comparisons (P<0.05). Kruskal-Wallis ANOVA with Tukey-Post-Hoc analysis of QK and expert scores identified a difference in the presence of ghost artifacts in images acquired with sequences A and B or B and C.


Statistical comparisons of QK scores identified a difference in the presence of ghost artifacts among the three sequences in full agreement with expert findings. This indicates that this kurtosis-based method can assess the variability of artifact presence in a stack without the need to process each image separately. In contrast to other methods, the proposed approach uses high order statistics (kurtosis) of background pixels to estimate ghost presence and is robust against (coil) bias due to the division with per-pixel mean image IA. Although further studies are needed, the proposed approach may be useful in readily assessing image quality in a clinical/research setting in an unbiased and fully-automated manner.

Authors’ Affiliations

Northwestern University
Northwestern University


© Tsaftaris et al; licensee BioMed Central Ltd. 2011

This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (<url>http://creativecommons.org/licenses/by/2.0</url>), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.