Development of a classifier for [18F]fluorodeoxyglucose extravasation severity using semi-quantitative readings from topically applied detectors

Overview

Development and evaluation of the extravasation severity classifier, that inputs topical detector readings, were performed with a retrospective human subject study. Subjects included those monitored with the topical detector system, during tracer uptake, and then scanned with FDG PET/CT (“Patient data” section). The topical detector system is detailed in the “Extravasation severity topical detector” section. Ground truth extravasation activity was measured from static PET scans using a volume of interest (VOI) approach  (“PET-based extravasation activity estimation” section). Due to the removal of topical detectors a median of 15 min before PET/CT, we utilized topical detector count rates extrapolated to the start of PET imaging (“Topical detector count rate extrapolation” section), enabling a direct comparison of static PET and topical detector readings. The extravasation severity classifier is described in “Semi-quantitative topical detector extravasation severity classifier” section and uses a regression of ground truth PET measurements against topical detector readings to produce semi-quantitative estimates of extravasation activity from the latter. Extravasation severity classification performance of the topical detector classifier was compared against a qualitative review by a radiologist and the ground truth PET values (“Radiological comparison study” section).

Patient data

The retrospective study examined patients whose FDG injections had been monitored with the topical detector system, from the time of injection until just prior to imaging, as part of their standard clinical workup for PET/CT. This study was approved by the Carilion Clinic Institutional Review Board (IRB), and the need for written informed consent was waived by the IRB. This retrospective study included a total of 24 subjects, with injection activity and uptake time of 320 ± 86 MBq and 68.0 ± 13.7 min (mean ± σ), respectively. Due to the use of de-identified datasets, patient gender and age were not available in this retrospective review. However, no exclusion criteria based on gender were utilized, and we only excluded vulnerable populations less than 18 or greater than 89 years old, for subject data collection. The study population contains subjects selected to represent a range of extravasation severity levels, identified through a qualitative assessment of topical detector TACs. Additionally, selected subjects had been imaged with the injection location (i.e., antecubital fossa, hand, etc.) positioned within the PET FoV.

The technologist utilized standard practice for selecting the injection location (e.g., preferentially inject in the arm contralateral to the primary cancer). When possible, subjects were injected with an infusion system (MEDRAD Intego, Bayer), with the infusion lasting ~ 1–2 min and included periods of saline flushing. In cases where the infusion system was not available, the dose was administered manually in ~ 1–10 s., followed by a saline flush. During the injection process and monitoring with the topical detector system, patients were positioned in an upright uptake chair with arms placed at their side on armrests. Subjects were asked to avoid excessive movement, but were allowed to move their arms as necessary during uptake.

Subjects were scanned on a Siemens Biograph mCT, with FlowMotion, or a Biograph TruePoint PET/CT (Siemens Medical Solutions USA, Inc.) and images reconstructed based on standard of care. Importantly, the Biograph mCT has a time-of-flight-capable PET camera, while the Biograph TruePoint does not. For both scanners, the imaging protocol began with a topogram, CT acquisition (120 kV and variable mA), and PET scan covering the subject from the skull base to the upper thighs. For the Biograph mCT, the PET acquisition utilized a continuous bed motion (table speed ranging from 0.7 to 1.6 mm/s., step-and-shoot equivalent of 180 and 80 s./bed, respectively), and images were reconstructed with time-of-flight ordered-subsets expectation maximization (OSEM) using 2 iterations and 21 subsets. For patients imaged on the Biograph TruePoint PET/CT, multiple bed positions with a step-and-shoot protocol were used to cover the subject with a scan duration of 3 min each (5 min for subjects above 113 kg), and images were reconstructed with OSEM using 4 iterations and 8 subsets. For both systems, reconstructed images were fully corrected for all factors impacting quantification (e.g., attenuation, scatter, etc.), sampled at 4.1 × 4.1 mm in axial slices, and smoothed in post-reconstruction with a 5-mm FWHM Gaussian filter. Slice thicknesses were 5.0 and 3.0 mm, for the Biograph mCT and Biograph TruePoint images, respectively.

Extravasation severity topical detector

An overview of the topical detector system, as well as sample datasets, is shown in Fig. 1. The system consists of two 511 keV gamma ray detectors. Complete specifications and performance measurements (e.g., count rate and energy linearity) of the system have been detailed by Knowland et al. [12]. Each detector is constructed of a monolithic 3 × 3 × 3 mm bismuth germanate (BGO) scintillator coupled to a silicon photomultiplier. No attenuative collimation is employed to restrict gamma ray incident angles. For readout electronics, a lower level energy discriminator of ~ 504 keV is applied and gamma ray flux (cps) is output for each detector at 1 Hz. Importantly, each detector measures and outputs photon count rates independently (i.e., no coincidence detection between detectors is utilized). Error in count rate linearity for a single topical detector for 18F was ≤ 1.5%, up to a true count rate of 80 kcps [12]. Both detectors are adhesively attached to the patient’s arms prior to radiotracer injection, with one detector (injection detector) placed proximal to the injection site and the other (reference detector) positioned in a matched location on the contralateral arm. TACs are then recorded continuously beginning immediately prior to tracer injection and during nearly the full uptake period. The main operating assumption is that the background signal from the body and arms is comparable in both detectors, such that the difference in TACs is primarily due to excess activity from an extravasation in the injection arm (Fig. 1C, D). We note that this operating assumption has been validated with a human subjects study comparing dynamic PET scans (ground truth) against topical detector readings [7]. In the absence of extravasations, TACs between injection and reference detectors matched closely. We provide additional support for this operating assumption here through a Monte Carlo simulation using a digital anthropomorphic phantom (see Additional file 1). TACs are corrected for tracer decay. No correction for solid angle, scattered photons, or photon attenuation is applied.

Fig. 1figure 1

Overview of the topical detector system used for classifying extravasation severity. A Single detector electronics that include a BGO scintillator coupled to a silicon photomultiplier and (B) placement of the two system detectors on the subject’s injection and contralateral arms; injection and reference detectors, respectively. C, D Representative TACs demonstrating signals from the injection and reference detectors, and their difference for subjects (C) with negligible and (D) moderate extravasation

Data processingPET-based extravasation activity estimation

Ground truth estimates of total extravasation activity were measured from PET images using a VOI-based approach adapted from that of Silva-Rodríquez et al. [13]. Noting the recorded injection location (e.g., right hand), both the tissue surrounding the injection (injection arm) and matched anatomical area of the contralateral arm (reference arm) were manually “painted” with initial VOIs on axial PET image slices using 3D Slicer [14]. An example of these VOIs is shown in Fig. 2. Initial VOIs were deliberately drawn to exceed the qualitative boundaries of both extravasation and normal physiological uptake. Total activity in the injection area (\(I_}}}\)), above what would be expected with normal physiological uptake alone, was computed as follows:

$$I_}}} = V_}} \left[ m_},i}} x_ - \frac}} }}}} }}\mathop \sum \limits_ m_},i}} x_ } \right]$$

(1)

where \(V_}}\) is the reconstructed image voxel volume, \(m_}}\) and \(m_}}\) are the binary masks for the injection and reference arms, respectively, \(x_\) is the PET image in units of Bq/ml, and \(V_}}\) and \(V_}}\) are total volumes for the injection and reference binary VOIs, respectively. Both \(m_}}\) and \(m_}}\) were produced from the original VOIs by imposing an SUV threshold of 0.2 to mitigate the impact of noisy background voxels.

Fig. 2figure 2

Representative VOI placement for the PET-based extravasation activity estimation method. VOIs were placed on the injection and contralateral arms, “Injection VOI” and “Reference VOI,” respectively

Topical detector count rate extrapolation

The difference of topical detector count rates was extrapolated from the time of detector removal to the start of PET imaging to enable a direct comparison against static ground truth PET extravasation activity estimates. Topical detectors were removed a median of 15 min before the start of PET imaging to allow time for the patient to void and for positioning in the PET scanner. Thus, besides the purpose of extrapolation, we do not use the dynamic information in topical detector TACs for extravasation severity classification in this study. The extrapolation method proceeded as follows: (1) take the difference of injection and reference arm TACs, (2) increase the extrapolation time start point past early segments, with large baseline transitions, utilizing a gradient-based approach, and (3) employ robust time-shifted decaying exponential function regression with the bisquare weights method and fit weights (\(W_ )\) calculated as:

where \(t_\) is the sample time, for sample i, and \(t_\) the fit start time. Nonlinear least squares regression was performed on up to 20 min of data and as little as 8 min, depending on the length of the TAC and the adjusted fit start time. We refer to the difference in topical detector count rates, extrapolated to the start of PET imaging, as \(\hat\). The performance of the extrapolation method was assessed using a total of 23 relatively long (time \(\ge\) 40 min) topical detector TACs taken from a subset of the patient datasets described in the “Patient data” section. TACs were synthetically clipped by 15 min and extrapolated to the artificial end time. Ground truth count rates were calculated from the difference of the median injection and reference TACs taken over 3 min.

Semi-quantitative topical detector extravasation severity classifier

A simple decision-tree-based multiclass extravasation severity classifier, inputting topical detector measurements, was derived from the PET ground truth extravasation activity values. For developing the classifier and assessing the general quantification of topical detector measurements, we performed the following steps: (1) topical detector difference readings were extrapolated to the start of PET imaging (\(\hat\) in the “Topical detector count rate extrapolation” section) and adjusted for radionuclide decay and positron fraction, and (2) PET image extravasation activities (\(I_}}}\) in Bq) were correlated against the extrapolated topical values (\(\hat\) in cps) with two different linear regression strategies, enabling estimation of total extravasation activity (\(\hat_}}}\)) from topical detector count rates alone. Equations for the least squares regression models were as follows:

$$\hat_},}}} = b_ \hat$$

(3)

$$\ln \left( _},}}} } \right) = b_ + b_ \ln \left( } \right)$$

(4)

where \(\hat_},}}}\) and \(\hat_},}}}\) are estimates of total extravasation activity from the two models, \(b_\) and \(b_\) are offset and slope parameters, and \(\ln ()\) is the natural logarithm. Thus, (4) is a linear regression on log-transformed PET image extravasation activities and extrapolated topical count rates. Ordinary least squares regression was used to estimate parameters for both models. For convenience, we refer to (3) and (4) as linear and log fits, respectively.

Normalizing the estimated total extravasation activity (\(\hat_}}}\)) by the decay compensated injected dose (\(A_}}}\)) produces the fraction of activity in the tracer extravasation relative to the total injected activity (\(\hat_}}} /A_}}}\)). A total of four classes was then used to stratify normalized extravasation activity as follows: (1) none (\(\hat_}}} /A_}}}\) < 1%), (2) minor (1% ≤ \(\hat_}}} /A_}}}\) < 5%), (3) moderate (5% ≤ \(\hat_}}} /A_}}}\) < 20%), and (4) severe (20% ≤ \(\hat_}}} /A_}}}\)). QIBA guidelines [6] specify \(\hat_}}} /A_}}}\) < 5% as minor and match the \(\hat_}}} /A_}}}\) moderate and severe classes given above. We rationalize the lowest threshold as follows: (1) static PET images provide no information on tracer extravasation resolution kinetics and (2) our prior experience with the qualitative impact of extravasations on PET images.

To evaluate the overall quantification of topical detector differences, the linear regression analysis of \(I_}}}\) versus \(\hat\) was performed on all human subjects data. For classification performance, a leave-one-out cross-validation was utilized to produce \(\hat_}}}\) values. We note that only the log fit, described by (4), was used to estimate total extravasation activity for assessing classification performance (see justification in the “Topical detector extravasation activity semi-quantification” section).

Radiological comparison study

The performance of the semi-quantitative topical detector extravasation severity classifier was compared against the qualitative assessment of a nuclear medicine radiologist (J.W.K) with more than 25 years of experience. Extravasation severity is currently evaluated qualitatively in the clinic, with no standardized methodology available. Thus, this approach represents the clinically utilized and preferred scheme of the associated nuclear medicine radiologist for classifying extravasations. PET images were ordered randomly with respect to patient ID, de-identified of patient information, and reviewed individually on a syngo.via workstation (Siemens Medical Solutions USA, Inc.). Images evaluated included maximum intensity projections (MIPs) and axial, coronal, and sagittal slices. The injection location (e.g., right hand) for each subject was accessible by the radiologist. Extravasation severity was scored on a scale of 0–3 as follows: 0 (none) indicated no visible evidence of radiotracer at or around the injection site, a score of 1 (minor) demonstrated small abnormal accumulation visible at or near the injection site that did not cause visible degradation of regional anatomy or image quality, 2 (moderate) showed more extensive visual evidence of extravasation at or near the injection site causing image distortion, blurring or loss of resolution of the regional anatomy or structures in proximity to the injection site, and 3 (severe) indicated a very large radiotracer extravasation that completely obscures regional anatomy such that no structures are recognizable. Additionally, in the arms down position, severe extravasations were those that degraded image quality of adjacent internal structures such as the liver, spleen, or kidneys and caused image windowing limitations due to scaling problems.

A binary classification analysis was performed by labeling all extravasations measured on PET with an uptake of \(I_}}} /A_}}}\) ≥ 5% as positives. For the topical detector classifier and radiological analysis, labels of moderate and severe were deemed positives. Performance was evaluated for this binary classification with the Matthews correlation coefficient (MCC). We also determined the optimal extravasation activity threshold (\(\hat_}}} /A_}}}\)), for topical detector measurements alone, to maximize binary diagnostic performance.

Statistical analysis

A percentile bootstrap method [15] (10,000 random samples) was utilized to estimate 95% confidence intervals of coefficients of determination from linear regression strategies. A Student’s t-test was used to test the significance between topical detector and radiological analysis binary classification results. Standard deviation was the result of a jackknifed standard error calculation. A P value of < 0.05 was deemed significant.

留言 (0)

沒有登入
gif