Paul J.C. Hughes1, Helen Marshall1, Felix C. Horn1, Guilhem J. Collier1, and Jim M. Wild1
1Academic Unit of Radiology, University of Sheffield, Sheffield, United Kingdom
Synopsis
Automating image analysis is key to accelerate quantitative image metric calculation and increase consistency between observers. This work presents a custom-built software to calculate percentage lung ventilated volume (%VV) from hyperpolarised gas and 1H MRI using spatial fuzzy c-means thresholding. The software developed reduced analysis time and user input resulting in significantly decreased interobserver variability when postprocessing image data.
Target Audience
Lung
MRI and image processing community
Background
Percent
ventilated volume (%VV) is a quantitative measure of lung ventilation derived
from hyperpolarised (HP) gas and
1H lung MRI often used in pulmonary
disease assessment for quantitative evaluation of early obstructive changes
1.
In order to calculate %VV, both HP gas ventilation images and proton anatomical
images must be segmented. Hard clustering methods have previously been used
2,3
for the segmentation of HP gas images with high success rates.
Developments in Fuzzy C-means segmentation algorithms have included the use of
spatial information in the membership function calculation
4 for
image segmentation providing an alternative to the k-means algorithms commonly
used.
Purpose
The
aims of this study were, (1) to develop a graphical user interface (GUI) for
semi-automatic %VV calculation using a spatial Fuzzy c-means (SFCM) algorithm,
and (2) to determine if the GUI would reduce interobserver variability when
compared with the existing workflow of manual thresholding and interpolation to
calculate %VV.
Methods
6
patient data sets were analysed after undergoing same-breath hyperpolarised
3He
MRI and
1H MRI at 1.5T
5 (GE HDx, Milwaukee, WI). Patients
were referred for scanning with a variety of pulmonary conditions including
asthma, COPD and CF, and ages ranged from 23 to 68 years (3 male, 3 female). A
manual method of %VV calculation was compared to a semi-automatic method using
the custom-built GUI:
Automated analysis
To apply
the SFCM method (figure 1) images were filtered using a bilateral filter
6
and segmented using a user-defined number of SFCM clusters. The user then chose
the most appropriate cluster representing the lung and this cluster was
transformed to a binary mask. Finally airways and any erroneous data were
excluded manually using ITK-SNAP
7, launched from within the GUI.
Manual
Data
was segmented using manual thresholding of ventilation images and manual
interpolation of the total lung volume defined from proton images overlaid with
ventilation images.
%VV calculation and analysis
Data was
analysed by two observers (O1 and O2) using both segmentation techniques to
calculate %VV as ventilated volume (from HP
3He images) divided by
total lung volume (from
1H images)
5. O1 and O2 had 6 and
1 years' experience of lung image segmentation respectively. Correlation,
Bland-Altman and Mann-Whitney analyses
were carried out.
Results
and Discussion
Interobserver correlation improved when using the SFCM method (r=0.99) when compared to using the manual method (r=0.93)
(figure 2). Furthermore the amount of bias between observers was reduced when
using the SFCM method (bias=1.10±2.87) when
compared to the manual method (bias=-6.96±6.55)
as seen in figure 3. Intermethod correlation for both observers was also high
(r=0.93 for O1 and r=0.87 for O2), although there is a relatively large
intermethod bias (bias=3.42±8.46 for O1 and bias=9.28±2.87 for O2). This bias is expected due to the way total lung volume is calculated
using information from the ventilation image in the manual method whereas the SFCM uses the anatomical image only, meaning some edge slices containing
minimal ventilation are excluded in the SFCM method but not in the manual method (e.g. due to partial-voluming
effects). No significant difference was found between the two observers’ SFCM measurements for all patient slices (n=229,
p=0.56) whilst there was a significant difference between the manual method
(n=229, p < 0.0001). The average time spent segmenting data sets was reduced
when using the SFCM method (22±5.5
mins) when compared to the manual method (38±7.5
mins). Furthermore the SFCM method is a user friendly
and simple way to segment lung MR images.
Conclusions
A GUI using SFCM has been developed and its
utility in calculating %VV has been demonstrated, with a significant reduction
in interobserver variability and analysis time in comparison with manual
thresholding and interpolation.
Acknowledgements
NIHR
and GlaxoSmithKline for fundingReferences
[1]
Woodhouse et al. J MAGN RESON IM, 2005, 21(4), 365-369; [2]
Zha et al PROC INTL SOC MAG RESON MED 23(2015), 1041;[3]
Kirby et al. ACAD RADIOL 2012, 19(2),
141-152; [4] Chuang et al. COMPUT MED IMAG GRAP, 2006, 30(1), 9-15;[5] Horn et al NMR BIOMED, DOI:10.1002/nbm.3187;[6]
Tomasi et al PROC IEEE ICCV 1998;[7] Yushkevich
et al. NEUROIMAGE, 2006, 1;31(3):1116-28;