Impact of time-of-flight PET on quantification accuracy and lesion detection in simultaneous ¹⁸F-choline PET/MRI for prostate cancer

Twenty men referred to our department for an ¹⁸F-choline PET/CT were prospectively enrolled in this study from December 2014 to January 2016.There was no further selection applied regarding patient inclusion. Exclusion criteria were (a) refusal to study participation or (b) claustrophobia or other contraindications for MRI (e.g., cardiac pacemakers). Informed consent was obtained from each patient prior to the study inclusion and to the PET/MRI. The local ethics committee approved the study under reference number KEK ZH-Nr 2013-0220.

All patients underwent a single injection of ¹⁸F-choline (mean dose ± standard deviation, 201 ± 4.7 MBq, range 195–214 MBq). Whole-body PET/MRI was performed after 59.3 min ± 4.9 min (mean ± SD) after injection. These PET/MRI study scans were acquired with a simultaneous TOF PET/MRI system (SIGNA PET/MR, GE Healthcare, Waukesha, WI, USA) used in previous studies at our department [33]: The scanner comprises a 3T wide-bore MR system with a TOF-PET detector ring installed between the body and gradient coils. The trans-axial and axial fields of view are 600 and 250 mm, respectively. The resolving time of the TOF detector is less than 400 ps [39].

The default acquisition protocol consisted of six bed positions (2 min per bed position), from the vertex of the skull to the mid-thighs. During PET scanning, MRI acquisitions for attenuation correction were performed, using in-phase and out-of-phase images to calculate water-only and fat-only images according to Dixon’s method [40, 41]. Additionally, the basic whole-body protocol and dedicated sequences to the pelvic region were applied. The MRI protocols and parameters are listed in Table 1.

An atlas-based attenuation correction was used for the head [42]. For the remaining body regions, air, lung, and soft tissue were segmented and a continuous fat-water based MRAC method was applied [43, 44].

3D-PET emission data was reconstructed with TOF and non-TOF for each patient with a fully three-dimensional iterative algorithm, which is part of the manufacturer-supplied standard scanner software (ordered subset expectation maximization (OSEM)-based VUE Point FX for TOF PET and VUE Point HD for non-TOF PET, GE Healthcare, Waukesha, WI, USA). Both algorithms include standard scatter, dead-time, random attenuation, and normalization correction as well as correction for the detector response using Sharper (GE Healthcare, Waukesha, WI, USA). Their only difference lies in the TOF information, which is ignored by the VUE Point HD algorithm. The PET images for both TOF and non-TOF were reconstructed using 2 iterations, 28 subsets, a 600-mm field of view, and a 256 × 256 image grid with 2.34 × 2.34 ×  2.78 mm³ voxels. For image space filtering, an in-plane Gaussian convolution kernel with a full width at half maximum of 4.0 mm was used, followed by a standard axial filter with a 3-slice kernel.

Reconstructed images were reviewed on a dedicated workstation (Advantage Workstation, Version 4.6, GE Healthcare, Milwaukee, WI, USA) for region of interest (ROI) analysis. Maximum standardized uptake values (SUV_max) was normalized to patient body weight. SUV_max was calculated for nine physiological structures using rectangle ROIs of 20 × 20 mm: skull, C7, Th12, L5, femoral head, pelvis, kidneys, ischioanal fossa, and M. gluteus. Average values were calculated for symmetrical structures. Lesions in the prostate, lymph node, and bone were assessed with size-adapted ROIs with knowledge of the reference standard. Additionally, size of lymph node metastases (maximal short axis) as well as minimum distance to osseous tissue (lymph node surface to cortical bone) was measured on T2-weighted MR images.

Four readers (readers 1 and 2, board-certified radiologist with 2 and 1 years of fellowship training in Nuclear Medicine; reader 3, 2 years of residency training in Nuclear Medicine; reader 4, 3 years of residency training in Radiology and 1 year of fellowship training in Nuclear Medicine) reviewed all reconstructed images on a dedicated review workstation. Readers investigated non-TOF PET and 1–3 months later TOF PET and were blinded to any other information than the patient bodyweight and injected dose (necessarily needed for the SUV calculations). Readers 1–4 evaluated the data for image quality, prostate lesions detection (or prostate bed if post-prostatectomy), lymph node metastases detection, and osseous and visceral metastases detection. Image quality was ranked from 4 (excellent) to 0 (non-diagnostic), and it was evaluated based on general quality, image sharpness, noise, and presence of artifacts (noise was evaluated in a scale from 0 (high noise) to 4 (low noise), artifacts from 0 (no artifacts) to 2 (severe artifacts), respectively). Readers 1 and 2 furthermore evaluated exact lymph node and osseous metastases site. An unblinded read out (board-certified Nuclear Medicine Physician and Radiologist), including all available histopathology and clinical/imaging follow-up data served as a reference standard. Details of the reference standard are shown in Table 2.

Difference of SUV_max (SUV_{max n-TOF} − SUV_{max TOF}) was compared using a two-sample paired t test and to achieve an alpha of 0.05, a conservative multiple comparisons correction of Bonferroni was applied (n = 12 different ROIs compared in each patient) [45]. Significant difference was defined at p < 0.05. Relative difference of SUV_max was defined as (SUV_{max n-TOF} − SUV_{max TOF})/SUV_{max TOF} and was compared using Bland Altman plots [46]. Results are presented as mean and relative standard deviation (RSD). Agreement among observers was evaluated using Fleiss’s kappa (κ) [47, 48] for all readers. Interpretation of κ was based on a classification provided by Landis and Koch: 0.0, poor; 0.0–0.20, slight; 0.21–0.40, fair; 0.41–0.60, moderate; 0.61–0.80, good; 0.81–1.00, almost-perfect reproducibility [49].

The Spearman’s rank correlation was used to investigate the effect of lymph node lesion size and the distance between lymph node lesions and bone on relative SUV_max difference. Lesion detection percentage was calculated for lymph node metastasis and distant metastasis. The Wilcoxon signed-rank test was used to compare visual image quality scores between TOF and non-TOF reconstructions. Neither sensitivity nor specificity measurements were calculated due to the heterogenous patient population and application of a heterogenous reference standard. All analyses were conducted in R 3.2.5 (The R Software Foundation, Vienna, Austria).

SUV_max was significantly lower on non-TOF images for all lesions, with an overall underestimation of − 17% (− 13% for prostate lesions, − 22% for lymph node metastases, and − 17% for osseous metastases, respectively). Typical distribution of relative differences of lesions is shown in Bland-Altman plots in Fig. 1c, d, and e. Spearman’s rank correlation showed a negative correlation for both lymph node lesion size and their distance to bone to relative SUV_max difference (ρ = 0.4, p value = 0.09 and ρ = 0.52, p value = 0.02, respectively) as shown in Fig. 3.

Detailed results of all four readers are given in Table 6. Reader 1 detected 84% (16/19) of lymph node metastases on TOF PET reconstructions (three missed metastases with histopathology reference) and 58% (11/19) on non-TOF PET reconstruction (five missed metastases with histopathology, three with imaging/clinical follow-up as reference, respectively). Reader 1 detected 93% (14/15) of osseous metastases on TOF PET and 73% (11/15) on non-TOF PET reconstructions (all missed metastases with imaging/clinical follow-up as reference). Reader 2 detected 84% (16/19) of lymph node metastases on TOF PET reconstructions (two missed metastases with histopathology reference, one with imaging/clinical follow-up as reference, respectively) and 73% (14/19) on non-TOF PET reconstruction (three missed metastases with histopathology, two with imaging/clinical follow-up as reference, respectively). Reader 2 detected 73% (11/15) of osseous metastases on TOF PET and 53% (8/15) on non-TOF PET reconstructions (all missed metastases with imaging/clinical follow-up as reference). Example of a missed lymph node metastasis is given in Fig. 4. The interreader agreement between readers 1 and 2 was good for localization of lymph node metastases for TOF and non-TOF (TOF kappa = 0.652, 95% confidence interval [CI] 0.506–0.798 and non-TOF kappa = 0611, CI 0.465–0.757). The interreader agreement between readers 1 and 2 was almost perfect for localization of osseous metastases for TOF (TOF kappa = 0.81, CI 0.631–0.989) and good for non-TOF (non-TOF kappa = 0.756. CI 0.577–0.935). Interreader agreement between all reader was good for local cancer detection for both TOF and non-TOF (TOF kappa = 0.762, CI 0.583–0.941 and non-TOF kappa = 0.696, CI 0.517–0.875), respectively. For lymph node metastasis detection, the agreement was moderate for both TOF and non-TOF (TOF kappa 0.501, CI 0.323–0.680 and non-TOF kappa = 0.457, CI 0.278–0.636). For osseous metastasis detection, the interreader agreement was good for TOF (TOF kappa = 0.636, CI 0.453–0.810) and moderate for non-TOF (non-TOF kappa = 0.600, CI 0.438–0.780).

General image quality, image sharpness, and image noise were rated statistically significantly superior (p value < 0.001 for general quality and sharpness, p value < 0.05 for noise) on TOF-PET reconstructions compared to non-TOF-PET reconstructions, there was no significant difference in the rating of image artifacts (p value = 0.25) (Fig. 5). Mean score (and SD) for TOF and non-TOF sequences were 3.0 (± 0.7) and 2.6 (± 0.7) for general quality, 3.1 (± 0.6) and 2.6 (± 0.7) for image sharpness, 2.7 (± 0.7), 2.3 (± 0.7) for image noise, and 0.4 (± 0.6) and 0.5 (± 0.6) for artifacts. Details of all readers are given in Additional file 1: Table S1.