Imaging protocol
Five patients with advanced colorectal cancer were included. Patients received 37.1 ± 0.9 MBq [89Zr] cetuximab within 2 h after administration of a therapeutic dose of 500 mg m−2 unlabelled cetuximab. Per patient, five PET/CT scans were acquired on a Gemini TF-64 PET/CT scanner (Philips Healthcare, Cleveland, OH, USA) at 1, 24, 48, 96 and 144 h post injection, respectively. PET data were normalized, corrected for decay, randoms, dead time, scatter and attenuation, and reconstructed using a time-of-flight list-mode ordered-subset expectation maximization reconstruction algorithm using an image matrix size of 144 × 144 and a voxel size of 4 × 4 × 4 mm3. In addition, for each time point, a 50-mAs ldCT scan was acquired for attenuation correction purposes. The corresponding ldCT images were reconstructed using an image matrix size of 512 × 512 and a voxel size of 1.17 × 1.17 × 5.00 mm3. The study was approved by the Medical Ethics Review Committee of the VU University Medical Center, and informed consent was obtained from each patient prior to inclusion in the study.
Delineation methods
Manual positioning
Low dose CT scans were first rebinned (with in-house developed software) using trilinear interpolation with a 4 × 4 × 4 mm3 voxel size in order to match matrix and voxel size of the PET images (Fig. 1a). Circular regions of interest (1.9-cm diameter) were positioned (in five slices per LV segment) in the bone marrow of all five LV segments by an experienced radiologist, and this was repeated for all five ldCT scans of each patient (Fig. 1c). Regions of interest were positioned in superimposed CT and PET so as to use complementary visual information, and five (cylindrical) VOIRM (diameter 1.9 cm, height 2 cm) of a total volume of 30 mL were extracted per ldCT scan. Subsequently, RM activity was determined by mapping the ldCT-derived manual VOIRM onto the corresponding PET scan.
Automatic delineation (active contour)
In a first initialization step, a region with a margin of about 1 cm was drawn around the five LV segments on the rebinned ldCT image (Fig. 1a). Next, this manual VOI was used to produce an LV binary mask that was applied to the ldCT image for extracting a coarse CT region that contained the LV component (CT1). The active contour model (see Additional file 1: Supplementary material) used in this study is primarily based on a work by Chan and Vese [8], with additionally incorporating a regularization term in the energy function that enables more robust segmentation of images with weak object boundaries [9]. Recently, the latter methodology was used by Sambuceti et al. [10] on CT data for extracting the whole bone marrow volume. The active contour model was developed in a MATLAB (MathWorks Inc., Natick, MA, USA) environment and was applied to CT1, where the LV bone contour (enclosing compact bone and bone marrow) was identified, providing a CT image of the compact bone and bone marrow (CT2). This LV bone contour was eroded to exclude compact bone; thus, a three-pixel layer was removed from the outer bone contour, resulting in a CT image containing only the intraosseous volume (CT3). The three-pixel layer was based on an educated guess after testing various pixel layers (from one to four pixels) on their impact in the final bone marrow volume and overlap with manually based ROI. Outer LV bone and intraosseous contours can be seen in Fig. 1b. The automatic delineation per subject (five CT scans) required 25 min on a 32-bit desktop PC using an Intel Core 2 Duo 2.8-GHz CPU with 3.2 GB of RAM, making it three times faster than the manual positioning of ROIs. Additionally, the automatic delineation allows the radiologist to perform in the meantime other clinical tasks.
Evaluation measures
The performance of each erosion kernel size (large, medium and small) was assessed by means of Dice similarity coefficient (DSC). This metric computes the volume overlap between the manual and automatic VOI as
$$ \mathrm{D}\mathrm{S}\mathrm{C}=\frac{2\left|A\cap M\right|}{\left|A\right|+\left|M\right|} $$
where A and M correspond to the automatic and manual VOI, respectively.
Organ dosimetry
After determining the mean activity concentration in a VOIRM at all five time points, RM time-activity curves were generated. Other organs were delineated semi-automatically to derive organ time-activity curves [11]. Cumulated activities were calculated as areas under the curves of RM and organ time-activity data by using the trapezoidal rule and assuming physical decay after the last measurement. The latter is a valid assumption, as in the supplementary material it is shown that the bone marrow activity concentration decreases with an effective half-life (73 h) that is shorter than the physical half-life of 89Zr (78.41 h). The residence time in the remainder of the body was calculated as the maximum possible residence time assuming physical decay only (no biological clearance) minus the sum of residence times of source organs. The organ residence times were scaled according to patient-specific weight data. Dose conversion factors (S values) were taken from OLINDA/EXM 1.1 software and were used for calculation of organ and RM absorbed doses [12]. Red bone marrow residence time data based on the plasma method were taken from Makris et al. [5], and the S
RM←RM, S
RM←RB and reference man/woman RM volume values used in the calculation of (self and total) RM absorbed dose were taken from OLINDA/EXM 1.1. S
RM←RM corresponds to the ‘red marrow to red marrow’ contribution, and S
RM←RB corresponds to the ‘remainder of the body to red marrow’ contribution. As this study focused on RM, only self and total (including contributions from source organs) RM doses will be reported.