The aim of this study was to assess different imaging and radiomics parameters for the distinction between microscopic liver tumor lesions and healthy liver volumes using 99mTc-labeled protein nanoparticle imaging. For the realistic modeling of clinical problems, we used a “healthy” population of mice that consisted of two groups, “control” and “obese.” A second, “tumorous” population of mice was included representing one “metastatic” and one “primary” liver tumor group.
Animals
Ethical permission was obtained from the institutional and national boards (see the “Ethics approval and consent to participate” section in the Declarations). Animals were fed ad libitum and maintained under controlled temperature, humidity, and light conditions.
Four groups of mice were examined in our experiment, a healthy population with two groups and a tumorous population with two other groups of mice. The first healthy group (control) contained different strains of mice (BALB/c, C3H, C57BL/6) to mimic population inhomogeneity with increased external validity (n = 2/breed, total n = 6, 32.4 ± 10.6 g, (mean ± SD)). Mice with higher body mass index—APP/PSEN1 transgenic mice and a C3H mouse strain with high body mass—(n = 8, 46.9 ± 14.5 g (mean ± SD) were used to model healthy but obese cases (obese group). Cancer-prone matrilin-2 knocked-out (KO) transgenic (MATN2) mice (n = 9, 38.5 ± 1.9 g (mean ± SD)) with chemically induced hepatocellular carcinoma were used as the primary liver cancer group. The fourth group was the liver metastatic group where severe combined immunodeficiency mice (n = 4; 21.3 ± 0.9 g (mean ± SD)) were inoculated with human melanoma cells (HT-168-M cell line, Avidin Ltd., Hungary) in the spleen. This model leads to fast liver metastasis formation. Mice were sourced from Innovo Ltd., Hungary.
Induction of tumorous models
The primary cancer group was induced by intraperitoneal diethylnitrosamine (Sigma-Aldrich, St. Louis, USA) injection in 15-day-old mice. The lack of matrillin-2 can increase the tumor proliferation in the liver [28]. The mice were investigated 4 months after the induction when liver tumors have reached the advanced multifocal status.
The metastatic group was induced by intrasplenic injection of a human melanoma HT168-M cell line (0.1 mL, ~ 5 × 105cells). This xenograft has high liver-colonizing capacity [29]. These mice were imaged only two months after the cell inoculation to mimic the early stage and the small size of metastatic lesions.
In vivo SPECT imaging protocol
Mice were anesthetized with isoflurane (3.5–4% induction and then reduced to 1.5–2% for maintenance of anesthesia) for the radiopharmaceutical administration and later again for imaging. Two hours before SPECT imaging, 87.2 ± 8.3 MBq (injected activity; mean ± SD) of 99mTc-protein nanoparticles, reconstituted from a clinically marketed human serum albumin-based kit for radiopharmaceuticals (Nano-Albumon® particles with < 200 nm size; Medi-Radiopharma Ltd., Budapest, Hungary; 2–5 GBq/mg specific activity), were administered intravenously. Animals were weighed prior to imaging.
SPECT (NanoSPECT/CT Silver Upgrade, Mediso Ltd., Budapest, Hungary) measurements were performed with multi-pinhole mouse collimators (pinhole diameter 0.7 mm). Abdominal SPECT scans were performed with 17–25 min scan times. The detection peak energy for 99mTc was set to 140 keV with a 20% wide symmetric energy window. The reconstruction of the raw SPECT data was performed with 0.33 mm isovoxel size while the field of view was centered to the abdominal region. All reconstructions were made by the instrument’s dedicated HiSPECT (Scivis GmbH, Göttingen, Germany) software. The reconstruction algorithm and settings may have an effect on the skewness and kurtosis values of voxel radioactivity distribution histograms. To minimize this source of error, we always applied the same reconstruction settings (“high” smoothing (33% Gauss), “high” reconstruction voxel size (0.33 μm), “low” iteration number (3 iterations) of the software settings). (In the primary group, we used the same raw data that were published before [27] for further evaluations.)
Histology
After imaging the animals were euthanized by pentobarbital sodium injection (400 mg/ml Euthasol, Medicus Partner Kft., Hungary). Autopsy was performed immediately and the liver was removed and stored in 4% formaldehyde-PBS solution. Afterwards, the presence of tumorous lesions was examined by histology in Hematoxylin-Eosin stained 10 μm thin sections.
Image and data analysis
After the reconstruction of image volumes, segmentation of the whole liver was made using the VivoQuant software (version 1.22., inviCRO, Boston, US) by two independent observers. At first, the liver region was visually cropped and an Otsu thresholding method was applied to obtain the reliable and automatic delineation of liver versus non-liver volumes [30]. The possible remnants of the spleen volume were then eliminated manually to get the final volume of interest (VOI).
These VOIs were compared to each other while two naïve observers were looking for liver cold spots. Four different image parameters (SUV, SUC, skewness, and kurtosis) were calculated from the segmented livers using MATLAB software (version 7.10.0., MathWorks, US) (the script and the calculated parameters are available in the Additional file 1) based on the following equations:
$$ \mathrm{SUV}=\frac{A_{\mathrm{liver}}\bullet m}{A_{\mathrm{total}}\bullet {V}_{\mathrm{liver}}}, $$
(1)
$$ SUC=\frac{A_{\mathrm{liver}}}{A_{\mathrm{total}}\bullet {V}_{\mathrm{liver}}}, $$
(2)
$$ \mathrm{skewness}=\frac{\frac{1}{n}\bullet \sum \limits_{i=1}^n{\left({x}_i-\overline{x}\right)}^3}{{\left(\sqrt{\frac{1}{n-1}\bullet \sum \limits_{i=1}^n{\left({x}_i-\overline{x}\right)}^2}\right)}^3}, $$
(3)
$$ \mathrm{excess}\ \mathrm{kurtosis}=\frac{\frac{1}{n}\bullet \sum \limits_{i=1}^n{\left({x}_i-\overline{x}\right)}^4}{{\left(\frac{1}{n-1}\bullet \sum \limits_{i=1}^n{\left({x}_i-\overline{x}\right)}^2\right)}^2}-3, $$
(4)
where Vliver is the radioactivity volume in cm3, Aliver is the summarized liver activity in kBq, Atotal is the administered radioactivity in kBq, and m is the mass of the animal in grams. In Eqs. 3 and 4. xi represents the radioactivity in each VOI voxel, x̅ is the arithmetic mean of xi values, and n is the number of voxels in the VOI.
Statistical calculations
Before the detailed analysis, the effect of manual segmentation process had to be verified. For this reason, we compared each of the resulting parameters based on two paired segmentations made by two independent observers. At first, difference plots (also known as Bland-Altman plot) were created for each imaging parameter, where the difference versus the mean of the repeated measures were plotted. Monotonic association (trend) between the difference and the mean of the repeated measures was examined with Spearman’s correlation tests. We preferred non-parametric methods due to the assumed non-normality and heteroscedasticity to describe the agreement between the two researchers. Therefore, Kendall W and the corresponding p values were calculated for each parameter. To check the systemic difference between the researchers a linear correlation was assumed. The slope and the intercept were determined based on this model. Hereinafter, we used the VOIs of the first evaluator.
Box-plots were created to comparing the differences of each parameter among the groups. Mann-Whitney U tests were performed to assess the differences of averages between tumorous and healthy populations. Multiplicity correction was not applied.
To judge the performance of the parameters’ classification between tumorous and healthy animal populations, empirical nonparametric receiver operating characteristics curves (ROC) were created. For plotting the ROC curve theoretical false and true positive rates (FPR and TPR) have to be estimated. As a crude estimation we made an assumption on FPR and TPR at each value in our dataset as a threshold for classification. In the case of SUV and SUC we defined cases as positive (assumed to be “tumorous”) if the value was smaller than the threshold. In the case of skewness and kurtosis we defined cases as positive if the value was larger than the threshold. The area under the curve (AUC) is estimated by summing the trapezoids enclosed by the points of the ROC curve.
Statistical calculations were made using R (Version 3.2.3. R Core Team, R: A Language and Environment for Statistical, Vienna, Austria, using vegan package for Kendall W). Difference plots were made by R using BlandAltmanLeh package. Box plots were created by Statistica 64 (Version 13., Dell Inc., Tulsa, US).