Introduction
The severity of chronic obstructive pulmonary disease (COPD) is clinically diagnosed
with spirometry by measuring forced expiratory volume in 1 s (FEV1). However, this
measurement does not give any information about the obstructed regions in the lung.
A better assessment of small airway diseases can be obtained with lung scintigraphy.[1 ] Ventilation/perfusion (V/Q) single-photon emission computed tomography (SPECT) has
also been incorporated for airway function diagnosis.[2 ]
For regional lung obstruction imaging, tomographic three-dimensional (3D) images obtained
in SPECT have a clear advantage over 2D projection images obtained in scintigraphy.
Technegas (Cyclomedica Ltd.) is used to assess lung ventilation function.[3 ] Technegas is a pseudo-gas, consisting of carbon nanoparticles suspended in argon
and labeled with 99mTc. The size of carbon nanoparticles is about 100–300 nm. This
allows for pulmonary ventilation imaging in nearly physiological conditions, since
nanoparticles can reach the alveolar space.[2 ]
As an alternative to SPECT, studies have reported the feasibility and implementation
of Gallgas-positron emission tomography (PET) scans for lung ventilation.[4 ],[5 ],[6 ],[7 ] Gallgas is a radiotracer made with the same carbon nanoparticles as Technegas but
labeled with 68Ga. Gallium-68 is an attractive PET radionuclide due to its short half-life
(~68 min), ease of production, availability, and compatibility with medical applications.[8 ],[9 ] Qualitative comparison studies between Technegas-SPECT and Gallgas-PET have been
reported,[10 ] where Gallgas PET shows a better distribution inside the lungs providing a better
visualization of ventilation heterogeneities. Other comparative studies have shown
that diagnoses with Technegas-SPECT and Gallgas-PET are compatible most of the time;
however, PET offers more confidence diagnosis in some cases.[6 ]
It is well known that PET has better spatial resolution than SPECT.[11 ] However, SPECT has the potential to become a more sophisticated technology through
image analysis techniques.[2 ] On the other hand, PET is gaining relevance in diagnose of pulmonary imaging[12 ] and will most likely continue to be used for this purpose in the future. For these
reasons, we considered a quantitative comparison between Gallgas-PET and Technegas-SPECT
to be of interest.
Previous authors have investigated the feasibility of selecting obstructed lung regions
in the lungs using textural features calculated from computed tomography (CT) images.[13 ] They found that a very good agreement with the segmentation obtained through texture
features' quantification and visual assessment done by experienced physicians.
We hypothesize that quantification in both techniques is relevant for diagnostic/prognostic
purposes. Moreover, advanced quantitative parameters such as texture features, measured
from ventilation PET and SPECT scans, are likely to show correlations to FEV1 as the
physiologic variable used to assess COPD severity.
Subjects and Methods
Patients
This study was approved by the Institutional Review Board and all participants signed
an informed consent form. For this retrospective study, five patients with appropriate
nonattenuation-corrected Gallgas-PET scans and available Technegas nonattenuation-corrected
SPECT scans were selected. Nonattenuation correction scans were used to compare PET
to SPECT acquisition mode. All these patients had moderate-to-severe COPD and underwent
Gallgas-PET scans at our center between September 2011 and July 2012. PET and SPECT
scans were performed within 4 days of each other. [[Table 1 ]] shows the details for the patients.{Table 1}
Table 1 Details of patient cohort
Radiopharmaceutical production
Technetium -99m was obtained from a 99 Mo/99mTc generator. For gallium-68, an eluate
from a 68Ge/68Ga generator was purified and concentrated by a QMA cartridge before
production of Gallgas nanoparticles. In both cases, carbon nanoparticles were produced
with a commercial device from Cyclomedica Ltd. Both radioisotopes were bound to carbon
particles in a pure argon atmosphere and a temperature of 2500°C. For Gallgas, carbon
nanoparticles were labeled with 68Ga as GaCl3. For Technegas, nanoparticles were labeled
with 99mTc as 99mTcO4-.
Imaging
An estimated mean inhaled activity of 70 MBq for Technegas and 27 MBq for Gallgas
was administered to the patients, instants before scanning. Ventilation SPECT was
performed acquiring 120 images around a 360° angle, with 15 s per step. Gallgas PET/CT
was performed using 2 or 3 bed positions, 3 min each, with a 64-slice multidetector
CT component. Each bed position comprises 46 slices. A bed overlap of 11 slices was
used (approximately 24%). Both images were reconstructed using an iterative OSEM algorithm.
PET images were corrected using time of flight.
Region of interest definition
The analyzed voxels from PET and SPECT scans were limited to the pulmonary region.
This region-of-interest (ROI) was segmented using the inspiration CT images and extrapolated
to PET and SPECT scans. All ROIs underwent visual inspection for further verification.
Trachea and main bronchus regions were removed manually from the ROI.
Image analysis
SPECT data were not corrected for attenuation, since it was acquired using a dedicated
SPECT camera. To compare both imaging techniques, we used the nonattenuation-corrected
PET data. SPECT scans were rigidly registered to PET using normalized mutual information
maximization criteria. After rigid registration, SPECT scans data grid sizes were
resampled to PET scans' data grids using nearest-neighbor interpolation method.
Volume segmentation
Threshold-based segmentation was performed by taking into account voxels with values
within a fixed range from SPECT and PET scan data. Pearson's correlation coefficients
between segmented volumes and FEV1 were calculated to find optimal segmentation thresholds.
Voxel histograms were calculated considering voxels inside the segmented pulmonary
region. For these voxels, an estimated value for background activity was calculated
using the mean value of voxels from the surrounding area outside the lungs. Voxels
with values lower than background were not considered for the histograms.
Texture features
Texture features were measured from PET and SPECT scans. To quantify these values,
we calculated parametric maps from scans using patches. A patch is defined as a cube
with N × N × N voxels (N-patch) centered in a single voxel. To obtain a parametric
map, each voxel is replaced by the value of a single feature calculated inside the
N-patch centered in that voxel. The final result is a 3D volume with the same size
as the scan data, each parametric map corresponds to a single texture feature. In
this case, we used a 5-patch size for calculations.
To assign a feature a value from a given ROI, the mean value of the corresponding
parametric map voxels inside this ROI is calculated. Voxels with neighboring patches
that fell outside the pulmonary region were not taken into account.
Fifty different first, second, and higher order texture features were calculated from
tomographic PET and SPECT scans.
First-order features: mean (1), median (2), variance (3), coefficient of variation
(4), skewness (5), kurtosis (6), energy (7), and entropy (8).[14 ]
Second-order features from gray level co-occurrence matrix: angular moment (9), contrast
(10), correlation (11), sum of square variance (12), inverse difference moment (13),
sum average (14), sum variance (15), sum entropy (16), entropy (17), difference variance
(18), difference entropy (19), information measure of correlation (20 and 21), maximum
correlation coefficient (22), maximal probability (23), diagonal moment (24), dissimilarity
(25), difference energy (26), inertia (27), inverse difference moment (28), sum energy
(29), cluster shade (30), and cluster prominence (31).[15 ]
High-order features using gray level run length matrix: short run emphasis (32), long
run emphasis (33), gray-level nonuniformity (34), run length nonuniformity (35), run
percentage (36), low gray-level run emphasis (37), high gray-level run emphasis (38),
short-run low gray-level emphasis (39), short-run high gray-level emphasis (40), long
run low gray-level emphasis (41), and long run high gray-level emphasis (42).[16 ]
High-order features using neighboring gray level dependence matrix: small number emphasis
(43), large number emphasis (44), number nonuniformity (45), second moment (46), and
entropy (47).[17 ]
High-order features using neighboring gray-tone difference matrix: coarseness (48),
contrast (49), and busyness (50).[18 ]
Correlations between texture feature segmentation and forced expiratory volume in
1 s
Using the previously described parametric maps, we used threshold-based segmentation
to divide voxels inside the lung region. We measured Pearson's correlation coefficients
between the total volume of the segmented regions and FEV1 volumes, for all 50 texture
features, after proper standardization of features' quantitative values. The highest
Pearson's correlation coefficient was used as criteria to find optimal segmentation
threshold using texture features.
Statistical comparison between texture features
Using the optimal segmentation thresholds measured from PET and SPECT, we selected
the voxels inside the obstruction regions common to both techniques. Thus, we took
into consideration voxels corresponding to obstructed regions for both PET and SPECT
scans. We compared the mean and variance values for each texture feature from all
these voxels taken from all five patients to compare SPECT and PET features.
All image analysis, quantification, and statistic tests were done using the software
MATLAB release 2017a (The MathWorks, Inc., Natick, Massachusetts, United States).
Results
Correlation coefficients between FEV1 and segmented volumes for several thresholds
are shown in [Figure 1 ]. Maximum correlation coefficient values are obtained for a threshold of 27% of maximum
uptake for PET and 31% of maximum uptake for SPECT scans, with Pearson's correlation
coefficients of 0.90 (P = 0.039) and 0.98 (P = 0.002), respectively.
Figure 1 Correlation coefficients between threshold-segmented volumes and forced expiratory
volume in 1 s
The histograms measured for each patient, from PET and SPECT scans, and the intensity
volume histograms (IVHs) are shown in [Figure 2 ]. IVH shows the volume in liters taking into account all the voxels with values above
a given threshold as a percentage of maximum uptake. [Figure 3 ] shows maximum intensity projections of PET and SPECT scans limited to the lung regions
for all five patients in the study.
Figure 2 Comparison between histograms of positron emission tomography and single-photon emission
computed tomography scans (first column) and intensity volume histograms (second column),
for each patient
Figure 3 Normalized voxel values positron emission tomography (top row) and single-photon
emission computed tomography (bottom row) maximum intensity projection scans. Each
column corresponds to scans for the same patient
Segmentation volumes using some of the 50 different texture features were found to
show high Pearson's correlation coefficients (>0.9) when compared to FEV1 as the physiologic
variable. These features, their corresponding threshold values for segmentation, together
with correlation coefficients and their P values are shown in [[Table 2 ]] and [[Table 3 ]] for Gallgas-PET and Technegas-SPECT, respectively.{Table 2}{Table 3}
Table 2 Texture features used for segmentation in positron emission tomography scans for
which resulting volumes show high correlation with forced expiratory volume in 1 s
Table 3 Texture features used for segmentation in singlephoton emission computed tomography
scans for which resulting volumes show high correlation with forced expiratory volume
in 1 s
Mean and standard deviation values of voxels inside obstructed lung regions were calculated
from parametric maps corresponding to these features, for both Technegas-SPECT and
Gallgas-PET scans. Comparison bar graphs for mean and standard deviation can be seen
in [Figure 4 ] and [Figure 5 ], respectively.
Figure 4 Mean values for the features with which segmentation volumes exhibit high correlation
with forced expiratory volume in 1 s, inside obstructed lung regions
Figure 5 Standard deviations for the features with which segmentation volumes exhibit high
correlation with forced expiratory volume in 1 s, inside obstructed lung regions
Discussion
FEV1 is the clinical quantitative variable generally used to assess COPD severity,
this is why we used these values to find a correlation with lung ventilation volumes
and thus obtain a significant threshold for segmentation in PET and SPECT scans. Optimal
thresholds (as percentages of maximum voxel value) used to segment ventilation volumes
were similar for PET and SPECT scans (27% and 31%, respectively). Previous studies
analyzed the relationship between V/Q ratio and FEV1.[19 ],[20 ] In this case, we analyzed correlations between ventilation scans and FEV1. A correlation
with a larger patient cohort between ventilation-segmented volumes and FEV1 could
reveal a link between tracer distribution and this physiological uptake. Visual inspection
shows that the uptake regions above this threshold correspond to what would be visually
assigned to the ventilated regions of the lungs. [Figure 6 ] shows the ventilated regions calculated using these thresholds in the same slice
of the same patient using PET and SPECT (images have been fused with CT for clarity).
Figure 6 single-photon emission computed tomography segmentation of ventilated region with
a threshold of 31% over the single positron emission computed tomography scan fused
with the computed tomography (a). Positron emission tomography scan segmentation of
ventilated region with a threshold of 27% over the positron emission tomography fused
with the computed tomography (b)
In an attempt to characterize tracer distribution, we plotted histograms and IVHs
from PET and SPECT scans and compared them for each one of our five patients. Comparisons
between histograms from PET and SPECT show similarities in some regions, but most
of them show differences for a given range of voxel values. These histograms are complemented
by the IVHs. Regions where histograms show more voxel counts in PET, together with
the higher level of the IVH curves translate the fact that there is a better peripheral
distribution of Gallgas when compared to Technegas. This is the other reason why we
used nonattenuation-corrected scans, to see more peripheral distribution of the tracer
in the lungs. Interestingly, even if three of the cases in this study showed the same
tendency, the other two showed an opposite, yet less marked, behavior. This suggests
that the radiopharmaceutical distribution could strongly depend on the cause of pulmonary
disease and the used radiotracer (in pseudogas form).
Texture features are useful when they can be related and used to assess certain aspects
of patient clinical data. In this case, we used textural features for threshold-based
segmentations using both SPECT and PET data in the lung tissues. Textural features
in the lungs have been studied as measured from CT scans.[13 ] To the best of our knowledge, there are no studies where textural features have
been quantified using SPECT or PET scans for lung ventilation. We chose a patch size
of 5 × 5 × 5 voxels, which by the voxel size accounts for 2.5 cm, as we considered
this to be large enough for the features not to be greatly affected by respiratory
motion. Nevertheless, it would be interesting to conduct sensitivity studies with
a larger patient sample and perform gated acquisitions. The segmented volumes we obtained
showed a good correlation to FEV1 volumes for 18 features in SPECT and 7 features
in PET. This suggests that there are intrinsic differences between the features measured
in the obstructed and ventilated uptake regions of the lungs, in both SPECT and PET,
and that textural features could play an important role in lung parenchyma segmentation
using these imaging modalities. Furthermore, as we can see in [Figure 4 ] and [Figure 5 ], some features exhibit similar statistical distributions in SPECT and PET scans.
We think that a standardized protocol could provide more robust texture feature quantification
in both techniques, as they seem to be consistent between these two scanning modalities.
Although these measurements should be taken with caution as we have only a few patients,
these results are encouraging to continue with similar trials.
Conclusion
To the best of our knowledge, this is the first study analyzing the differences between
PET and SPECT using quantitative metrics other than standardized uptake value and
V/Q. Taking this small cohort of patients into consideration, results suggest that
tracer distributions could strongly depend on the cause of the disease. Segmentation
using some texture features quantified from SPECT and PET was found to have a better
correlation to physiological variable FEV1, motivating to continue the research in
this field. Studies including a larger cohort of patients are necessary to have statistically
significant results.
