Two different PET/CT scanners were used, the Discovery-IQ (DIQ) and the Discovery-MI (DMI) (General Electric Healthcare, Chicago, IL, USA), both with a 5-ring configuration leading to an equivalent axial field of view (FOV). The technical specifications of the devices are detailed in Table 1. Considering the size and composition of the crystals in each system, we expect to observe differences in terms of detection limits and semi-quantification. The size of the crystals is a very important design parameter, as it has a direct impact on the spatial resolution of the PET systems. The DMI has detection crystals with thinner dimensions. In addition, the crystal used in the manufacture of the DMI is lutetium–yttrium-oxyorthosilicate (LYSO). Due to the properties of LYSO, it enables additional data correction using time of flight (TOF) information, resulting in an enhanced signal-to-noise ratio in the resulting PET images. These two factors suggest that the DMI should have the capability to detect more targets than the DIQ.

We designed a clinical study, IQversusMI (NCT03956459, May 20th, 2019), a monocentric prospective paired study designed to assess the relevance of ISL for evaluating PET performances from patient data. This study was conducted at the Institut Universitaire du Cancer Toulouse Oncopole (IUCT-O; Toulouse, France).

Eligible patient had cancer indication for 18F-FDG PET according to current clinical practice standards and were able to maintain a strict supine position on two occasions. They were ages > 18 years and had an Eastern Oncology Group Performance status (ECOG) of 0 or 1 and Karnofsky index > 70. The ECOG or Karnofsky Index assesses the overall health status of a patient. Conducting clinical research on patients in a compromised health state, particularly when there is no potential benefit for the patient, as is the situation in this study, is deemed unethical.

Key exclusion criteria were unbalanced diabetic patient and patients with a formal contraindication to certain imaging examinations (severe claustrophobia, heart valve, pacemaker, etc.). All patients provided written informed consent.

24 patients were stratified according to their body mass index (BMI) and the number of synthetic lesions to be inserted. Among the 24 patients enrolled, 12 had a BMI below or equal to 25 (BMI ≤ 25), while the other half were strictly above 25 (BMI > 25). In addition, both BMI cohorts were divided into three sub-cohorts in which the tumour burden varied. The detail of the stratification adopted for this clinical study is illustrated in Fig. 1.

Stratification depending on BMI and tumour burden (i.e. ISL) of patients enrolled in the IQversusMI clinical trial

Between July 2019 and February 2021, 24 patients were enrolled in this prospective monocentric study. In Table 2, we summarized demographic and baseline characteristics.

Patients underwent two consecutive PET/CT scans corresponding to different scanners, the DIQ and the DMI. The second examination was performed without additional injection of radiotracer and with a low-dose protocol CT. We defined anatomical locations identified on the consecutive exams of the same patient. Synthetic lesions were modelled and simulated on the PET images. An overview of the whole process is shown in Fig. 2.

Entire workflow of IQversusMI clinical trial from patient recruitment to the completion of reading sessions by physicians. The figure illustrates an OligoM case with three inserted lesions

Consecutive acquisitions were performed with an uptake time around 60 min and 85 min, respectively, after a single injection of 18F-FDG (2 MBq/kg). The order in which the examinations were performed was balanced in order to avoid the introduction of a bias due to the biodistribution of the radiotracer, which would differ at the time of each acquisition.

Acquisitions performed in list-mode were used to generate raw data with different durations in order to obtain similar counts per exam for each dataset. The second exam lasted longer to compensate for the radiotracer decay and to obtain equivalent statistics for both exams. These raw data were then reconstructed into PET images using Bayesian Penalized Likelihood (BPL) for the DIQ and with TOF additionally for the DMI. PET acquisition and reconstruction parameters are detailed in Table 3.

CT technologies composing DIQ and DMI systems are very similar. In clinical practice, we are using the same CT protocols for both PET/CT systems. CT acquisition parameters for the first scan are available in Table 4.

In addition, to minimize unnecessary radiological exposure, the second CT acquisition used a protocol with adjusted parameters for focusing specifically at attenuation correction, commonly known as CT-Based Attenuation Correction (CTAC). CTAC series were used for CT visualization during the reading session.

After anonymization, PET and CT images were then transferred to a research workstation for the ISL as described elsewhere [12]. We used a dedicated high-performance research workstation Z8 (Hewlett-Packard, Palo Alto, CA, USA) where the modelling, simulation and reconstruction were performed on a reconstruction research toolbox (Duetto v02.13, General Electric Healthcare, Chicago, IL, USA) and executed with MATLAB R2018b (The MathWorks Inc., Natick, MA, USA).

Synthetic lesions were inserted at the same anatomical location into the two consecutive exams of the same patient. Lesions were inserted in key organs characterized either by specific densities (lungs, bone, liver) or in the vicinity of fixed (bladder, kidney, heart) or mobile (diaphragm) organs.

We determined the same anatomical location and measured the mean activity concentration (AC) (kBq/mL) for each PET system. It was done to ensure close level of AC considering the same anatomical spot on consecutive exams of the same patient. Then, we chose the size in diameter and opted for a specific contrast, which was applied to the synthetic objects. The contrast was calculated using formula (1). We assessed the AC in the background by using a 2 cm3 spherical volume of interest (VOI) positioned on the insertion site prior to the simulation. We established the AC of the lesion based on the desired contrast.

$$}= \frac}\left(}\right)-}(}))}}(})}$$

(1)

Table 5 shows the characteristics of synthetic targets in terms of shape, size, contrast, AC and anatomical location.

During the study, volumes, contrasts and anatomical locations of synthetic lesions were not equally distributed between the two BMI groups. The reconstructed images were then imported and stored on a dedicated interpretation workstation, the AWServer client console (General Electric Healthcare, Chicago, IL, USA), for the reading sessions using the image interpretation software PETVCAR®.

Two nuclear medicine physicians that were unaware of the pathological indication evaluated the images by reporting the detected lesions and then measuring the semi-quantitative values of glycolytic activity with the standardized uptake value (SUVmax, SUVmean and SUVpeak). The images were assessed in a random order, and the observers were not aware of the PET system on which the examination was performed (DMI or DIQ) or of the number of inserted synthetic lesions to be identified. The observers compared their interpretations and came to a consensus.

The primary objective of this trial was to assess the relevance of ISL for evaluating PET performances from patient data at the lesion level. Sample size was computed for a lesion-based analysis [17, 18]. Based on the outcomes of clinical investigations, the sizes and contrasts considered for the synthetic targets and our extensive experience with these systems, we assume that the DMI should be able to detect at least 1.7 times more targets in contrast to the DIQ.

Assuming a 20% probability of agreement, 58 lesions per BMI group are required to demonstrate this difference with 80% power and a two-sided alpha of 5%. Knowing that a maximum of 15 lesions should be simulated per examination and that 58 lesions are required for each BMI group, we enrolled 12 patients per group according to the simulation strategy.

For each group of BMI (≤ 25 and > 25), lesions are simulated as follows: 4 patients without synthetic lesions (M0), 4 patients with 1 to 5 synthetic lesions (OligoM) for a total of 15 lesions and 4 patients with more than 10 synthetic lesions (MultiM) for a total of 45 lesions. The stratification of the clinical trial is illustrated in Fig. 1.

fPET/CT system and the corresponding RTPR was determined (RTPR) for synthetic lesion defined by the ratio of the detection rate of synthetic lesion by DMI and DIQ. The detection rate of synthetic lesion is defined by the ratio of the number of lesion detected by the PET and the total number of synthetic lesion. The RTPR for natural lesion was evaluated in a similar way. We assumed that the total number of natural lesions was equal to the number of natural lesions reported by at least one of the readers. Using this information, the detection rates for natural lesions were calculated for each PET/CT system and the corresponding RTPR was determined.

For assessing synthetic lesion inter-reader variability, we calculated inter-observer agreement (IOA). IOA was assessed by the concordance rate between the 2 readers. The concordance was estimated with 95% confidence interval (binomial exact). We considered an adequate inter-reader variability for IOA above 80%.

Additionally, by calculating the mean relative differences (RD) and standard deviations (SD) of lesions SUV metrics reported on the DIQ and DMI, we contrasted the semi-quantitation for synthetic and natural lesions. We determined the RD using formula (2). We calculated the SD from the distribution of the RD index.

$$}(}/})= \frac}}_}}-}}_}}\right)}}}_}}} \times 100$$

(2)