Image acquisition was performed with the micro-CT scanner of the trimodal PET/SPECT/CT Albira ARS preclinical system (Bruker, Spain). Micro-CT images were acquired with 45 kV, 0.8 mA, and 400 projections, according to a previously optimized imaging protocol for subtracted CE micro-CT imaging with this scanner [15]. This protocol involved the acquisition of two images: a pre-contrast image (i.e., the baseline image) and a second image after or during the administration of the contrast agent (i.e., the contrast-enhanced (CE) image). The baseline image was then subtracted from the CE image to yield the subtracted CE image, which was then parameterized to units of CI with a calibration function to yield the subtracted CI image. The total radiation dose to water measured at the isocenter for this imaging protocol was 680 mGy [15].

Image reconstruction was performed with either the conventional vendor’s reconstruction algorithm (FBP-based) or with a simultaneous iterative reconstruction technique (SIRT) algorithm implemented in-house with the Matlab R2018b (The MathWorks Inc., Natick, MA, USA) ASTRA toolbox. The SIRT algorithm was previously validated and optimized [22]. The number of iterations for the SIRT algorithm was optimized in this study as a trade-off between noise and spatial resolution, as described in the Supplementary Appendix and briefly summarized here. Increasing the number of iterations increased the noise and improved the spatial resolution in SIRT reconstructed images, as shown in Supplementary Figure S1. A range of 85–180 iterations were evaluated, and 180 iterations were selected since this number produced SIRT images with the highest spatial resolution among the evaluated number of iterations. An adverse consequence of selecting a high number of iterations, however, is that a higher noise content would be observed in the SIRT images used in this study, compared to the FBP images. FBP reconstructed images had a matrix size of 560 × 560x516 and a pixel size of 0.125 mm; SIRT reconstructed images had a matrix size of 750 × 750x657 and a pixel size of 0.1 mm. Images were calibrated to Hounsfield units (HU) using the average attenuation value of water for each reconstruction algorithm, which was obtained from images of a water phantom; a transverse view of this phantom is shown in Fig. 1.

(a) Water phantom, (b) acrylic semi-cylinder phantom, and (c) calibrated iodinated phantoms

A BF is a spatial-domain non-linear function designed to reduce noise while preserving the small structures and edges [23]. Its optimal parameters are related to the noise and spatial resolution of the image to be filtered, which means that they are unique for a given imaging protocol and reconstruction algorithm. In this work, a specific 3D BF was used for micro-CT images reconstructed with SIRT or FBP algorithms. BFs were implemented in MATLAB R2018b (The MathWorks Inc., Natick, MA, USA); the details of their implementation have been described previously [20].

Image quality was assessed with the noise power spectrum (NPS), the modulation transfer function (MTF), and the contrast-to-noise ratio (CNR) in four scenarios: FBP reconstructed images, SIRT reconstructed images, filtered FBP images (fFBP), and filtered SIRT images (fSIRT). All images were reconstructed with the FBP and SIRT algorithms. The fFBP and fSIRT images were obtained after applying the corresponding BF to the reconstructed images. The 2D and 1D NPS were evaluated from images of a water phantom, and the 1D MTF was evaluated in the transverse plane from images of an acrylic semi-cylinder phantom, following guidelines for the assessment of image quality in CT scanners [24, 25]. Figure 1 shows the water and acrylic phantoms.

Calibrated iodinated phantoms were used to assess the CNR[26]. These phantoms consisted of a solid epoxy material with CI values of 0, 0.5, 1.0 and 3.0 mg I/ml, as shown in Fig. 1. One image of each iodinated phantom was acquired in a separate and consecutive manner with the same acquisition parameters, as defined above. An affine registration was performed between each image and the image of the 0 mg I/ml phantom [27], which was considered the baseline image, and then the baseline image was subtracted from the images of the other iodinated phantoms to yield the subtracted CE images.

The contrast was quantified from the subtracted CE images as the difference between each iodinated phantom and the 0 mg I/ml phantom. The noise was evaluated as the standard deviation of the mean value measured in the subtracted CE image of the 0 mg I/ml phantom. The CNR was estimated as the contrast of each iodinated phantom divided by the noise. The NPS, MTF, and CNR were quantified using MATLAB R2018b (The MathWorks Inc., Natick, MA, USA).

The subtracted CE images were converted to CI using an appropriate calibration function for each reconstruction algorithm, to yield the CI images. The calibration functions (CI vs. HU measured in the subtracted CE images of the calibrated iodinated phantoms) were CI = 0.022*CE + 0.298 (R2 = 0.99) and CI = 0.023*CE – 0.095 (R2 = 0.99), for FBP and SIRT images, respectively.

Three repeated measurements of CI were performed for each nominal CI value. A separate set of images of the iodinated phantoms was acquired in a second experiment to assess the precision with a test–retest approach; CI values were measured in this set of images to yield the replicate measurements.

The linear relationship between the measured and the nominal CI values were assessed for the FBP, SIRT, fFBP, and fSIRT images [10]. Plots of the replicate measurements were obtained (measured vs. nominal CI); 2nd and 1st order polynomials were fitted to the data, and linearity was supported when the β2 coefficient of the 2nd order term of the 2nd order fitted polynomial was small (β2 < 0.5), and the β1 coefficient of the 1st order term of the 1st order fitted polynomial was close to one (0.95 < β1 < 1.05) and R2 > 0.9. The CI accuracy was assessed with the bias [10], which was determined as the difference between the measured value and the nominal value; the bias was plotted against the nominal value.

The precision of CI was estimated from repeatability and reproducibility metrics [10]. Repeatability was assessed with the within-subject standard deviation (wSD = standard deviation of the replicate measurements for each nominal CI value), the within-subject coefficient of variation (wCV = wSD/mean), and the repeatability coefficient (RC = 2.77wSD). The reproducibility of CI was evaluated with the correlation coefficient for the following comparisons: FBP vs. SIRT, FBP vs. fFBP, SIRT vs. fSIRT, and fFBP vs. fSIRT.

All experimental procedures with the animals were reviewed and approved by the Ethics Committee and the Institutional Animal Care and Use Committee of the Instituto Nacional de Cancerología, Mexico, where all the experiments took place; approval number: (018/051/IBI) (CEI/1294/18). The in vivo evaluation was performed on a virgin female Sprague–Dawley rat with chemically-induced mammary cancer. Mammary lesions were chemically induced with dimethylbenz[a]anthracene (DMBA) [28]. The animal was kept in a pathogen-free environment and fed with autoclaved food and water ad libitum. A single intragastric dose of 20 mg/ml DMBA (Sigma) dissolved in 1 ml of sunflower oil was administered to the animal (7-week-old), after a previous intraperitoneal injection of ketamine and xylazine (30 and 6 mg/kg body weight, respectively) [28]. Imaging was performed after tumor detection, which occurred 10 weeks after the inoculation of DMBA. For image acquisition, the animal was anesthetized with isoflurane (3% in 100% oxygen). A baseline image was acquired; then, a CE image was acquired during continuous infusion of a clinical contrast agent (Omnipaque 300, GE Healthcare,Wauwatosa, WI, USA; average dose = 2.4 mg of iodine/g of body weight (b.w.), infusion rate = 0.5 mL/min), via a catheter placed in the right external jugular vein of the animal. No gating (cardiac or respiratory) was used during image acquisition. Images were reconstructed, filtered, registered, and subtracted as described and converted to CI values. Mean CI and its standard deviation were quantified in FBP, SIRT, fFBP, and fSIRT images within spherical volumes of interest (VOIs) with AMIDE software [29] for several tissues. VOIs were placed in the left ventricle (LV, 3 mm diameter), abdominal aorta (0.7 mm diameter), liver (3 mm diameter), tumor (2 mm diameter), and muscle (2 mm diameter). A CNR related to muscle (CNRmuscle) was obtained for each tissue; in this case, the contrast was evaluated as the difference between CI within each tissue and CI within the muscle, and the noise was defined as the standard deviation of the mean value of CI within the muscle.

GraphPad Prism 6 (GraphPad Software, Inc., San Diego, CA, USA) was used to perform all statistical analyzes. A Shapiro–Wilk test was used to assess the normality of the data. Data were compared in the following pairs: FBP vs. SIRT, FBP vs. fFBP, SIRT vs. fSIRT, and fFBP vs. fSIRT. Normally distributed data were compared with a one-way analysis of variance (ANOVA) test, followed by Bonferroni’s test for multiple comparisons (namely, CNR, bias, RC, CNRmuscle). Non-parametric data were compared with the Friedman test and Dunn’s multiple comparison test (namely, NPS, MTF, CI in the in vivo evaluation). Pearson correlation coefficient was used to evaluate the reproducibility of CI. An adjusted p-value less than 0.05 was considered as statistically significant.