On the optimal z-score threshold for SISCOM analysis to localize the ictal onset zone

Patients were selected from an existing local database of epilepsy patients who underwent either temporal or extratemporal lobe resection from 1998 to 2014. Patients were included in this study if they (1) had undergone both ictal and interictal SPECT-imaging before surgery, (2) had a volumetric preoperative MRI, and (3) were completely seizure free in the postoperative period for at least 1 year after surgery or had less than three seizures in the postoperative period (less than 6 months after surgery) after which they were seizure free for at least 1 year [17]. In all patients, resective surgery strategy and extent of resection were determined by consensus at a patient management meeting with integration of all available clinical information and all paraclinical and technical investigations including neuropsychological evaluation, EEG, structural MRI, language or motor fMRI, FDG PET, and SISCOM. During these meetings, the routinely produced SISCOM with a z-score threshold of 2 had been used for clinical interpretation and decision making.

Patients were hospitalized and monitored by video and EEG to direct the ictal injection. ^99mTc-ECD (ethylcysteinate dimer) was prepared by kit labeling and injected by epilepsy-trained nurses immediately after clinical seizure onset [18, 19]. Timing of injection was determined on retrospective video and EEG review. Seizure onset was defined as the earliest ictal EEG or clinical evidence of seizure activity. The median injection time was 20.5 s (range 2–116 s), with 73% of patients having injections within 30 s. The median seizure duration was 79 s (range 16–389 s). Three injections in two patients were given post-ictally (two injections in one patient with extratemporal lobe epilepsy, one injection in one patient with temporal lobe epilepsy). In the group of patients with extratemporal lobe epilepsy, the median injection time was 19.5 s (range 1–64 s) with a median seizure duration of 47.5 s (range 16–219 s).

For the interictal study, the patient underwent 5 min of seizure-free EEG monitoring before injection and this monitoring was continued for at least 3 min afterwards. The interictal scan was started 20 min after tracer injection, while ictal scans were acquired between 30 min and 4 h after injection.

SPECT images were acquired using a triple-head Triad XLT (Trionix, Misouri, USA; from 1998 until 2007; protocol one: 120 steps × variable number of seconds (depending on the time between dose calibration and acquisition, 128 × 64 matrix) or a triple-head IRIX Prism (Philips Medical Systems, Cleveland, USA; from 2007 until 2014; protocol two: 120 steps × 15 s, 128 × 128 matrix). Reconstruction of the images was performed using filtered back projection (FBP) with a Butterworth filter at cutoff value of 0.87 cycles/cm and order 7 (protocol one) or using ordered-subset expectation maximization (OSEM) iterative reconstruction with six iterations, eight subsets (protocol two) without postsmoothing. The acquisition and reconstruction was always the same for the ictal and interictal SPECT scans of the same patient.

Ictal and interictal SPECT studies were co-registered using an automatic registration algorithm based on mutual information using the registration module of Statistical Parametric Mapping (SPM version 8; Wellcome Trust Centre for Neuroimaging, London, UK), implemented in Matlab (R2012; The MathWorks Inc., MA, USA) [20, 21]. The ictal and spatially co-registered interictal images were then normalized for global brain counts within each scan. The transformed, normalized interictal images were subtracted from the normalized ictal image to create an image where the value for each pixel represents the intensity difference between the two data sets. The difference image was smoothed using a 3D-Gaussian smoothing kernel (full width at half maximum = 15 mm) and transformed into a z-score using the mean and standard deviation (SD) of the differences in all brain voxels [8, 11, 20]. The average of both SPECT images was co-registered to the structural MRI, and this transformation was applied to the z-score image. The thresholded z-score map was then co-visualized onto the co-registered preoperative MRI using MRICRO software (Georgia Institute of Technology, Atlanta, GA, USA) for detailed anatomical localization.

Image readout was performed by two experienced nuclear medicine physicians with respectively 10 and 20 years of experience in SISCOM evaluation. Readers were blinded to all clinical information. To score, the brain cortex was divided into 10 volumes of interest (VOI; left and right frontal, temporal, perirolandic, and parietal and occipital cortex), as described previously [6, 16]. The readers were asked to assign, if possible, the IOZ from each individual SISCOM study to one or more of the 10 cerebral VOIs. IOZ was defined as the cortical area with the greatest subtraction values. When the cluster with the highest z-score value was located in the occipital lobes, this cluster was disregarded if other significant clusters in the brain were present, since these clusters in the occipital lobe are far more likely to represent a different activation of the visual cortex during the ictal (eyes open) compared to interictal (eyes closed) injection than an IOZ. The two sets of data with z-score thresholds of 1.5 and 2 were presented in a randomized order (the reader was only aware of which z-score threshold data belonged to). We only investigated thresholds of 1.5 and 2.0 since in the study by Newey et al. and in our clinical experience, z-score thresholds of 1.0 give rise to too many false positive clusters.

Readers were also asked to give a confidence level to their localizing score (one = not confident at all to five = very confident). The final localization was based on agreement between the two readers by a consensus read, in case of discrepancy after the first read. In all cases, a consensus agreement between the two readers was found. The study was considered non-localizing if no cluster was found that could fit the IOZ.

All statistical analyses were performed using SPSS v24 (IBM Business Analytics, Chicago, IL, USA). Significance was determined at the p < 0.05 level. With the surgical ground truth as reference, detection ratio was calculated for both thresholds and confidence levels were compared using Wilcoxon signed ranks test. Furthermore, the kappa (κ) coefficient and its confidence (SD) were calculated to analyze interrater variability in SPSS.

Eighty patients (45 male (56%), 35 female (44%)) met the inclusion criteria. Of those, 12 patients underwent two ictal SPECTS, so a total of 92 ictal SPECTs were available for analysis. Demographic and clinical characteristics are shown in Table 1. The mean age at surgery was 38.5 years (SD 11.4 years; range 14–59 years). The preoperative mean seizure burden was 21 seizures/month with 40 patients (50%) having a history of focal to bilateral tonic-clonic seizures. Sixty-one patients had temporal lobe epilepsy (76%), and 19 had extratemporal lobe epilepsy. In patients with extratemporal lobe epilepsy, IOZ was localized in the frontal lobe in 11 patients, in the parietal lobe in 5 patients, in the occipital lobe in 2 patients, and in the insula in 1 patient. Fifty-six (70%) patients had a temporal lobe resection, five patients (6.3%) had a temporal lesionectomy, one patient (1.3%) had temporal radiosurgery, and 11 patients (13.8%) had a lesionectomy in the frontal lobe, two patients (2.5%) in the occipital lobe, and five patients (6.3%) in the parietal lobe. Histopathological analysis was available in 78 patients and showed hippocampal sclerosis in 46 patients, focal cortical dysplasia in 25 patients, dysembryogenic neuro-endocrine tumor (DNET) in 4 patients, cavernous hemangioma or no abnormality in 3 patients, dual pathology or reactive astrogliosis in 2 patients, and ganglioglioma, epidermoid cyst, meningioma, oligodendroglioma, neurocytoma, or unclear findings in 1 patient. After surgery, all patients were seizure free for at least 1 year. In seven patients, one to three seizures were present in the immediate postoperative period but they were seizure free afterwards for the rest of their follow-up, which was at least 1 year.

At z-score threshold of 1.5, 1/92 scans was considered non-localizing vs. 4/92 scans at z-score threshold of 2.

At a z-score threshold of 1.5, 70% (64/92) of the studies were correctly localized by the consensus read. This value increased to 72% (66/92) at a z-score threshold of 2 (p = 0.17). An example of a typical patient is shown in Fig. 1. The rate of correct localization was significantly higher in patients with temporal versus extratemporal epilepsy at a z-score threshold of 1.5 (81 vs. 40%, p < 0.001) as well as at a z-score threshold of 2 (81 vs. 48%, p = 0.004). Within each group, however, there was no significant difference in rate of correct localization between a z-score threshold of 2 and 1.5.

Consensus read was necessary in 23 of the 92 SISCOMs analyzed with a z-score threshold of 1.5 and in 18 of the 92 SISCOMs analyzed with a z-score threshold of 2. Analysis of interrater reliability analysis demonstrated no significant differences in the correctness of the IOZ localization between a z-score threshold of 1.5 versus a z-score threshold of 2 for any reviewer (reader one: 63/92 at threshold 1.5 vs. 64/92 at threshold 2; reader two: 64/92 vs. 66/92). However, the reviewer with 20 years of experience had a significantly higher confidence level using a z-score of 2 vs. 1.5 (4.2 ± 1.1 vs. 4.0 ± 1.0, p = 0.002), in contrast to the reviewer with 10 years of experience (4.3 ± 1.0 vs. 4.2 ± 1.1, p = 0.39). The agreement in localization between the two reviewers was moderate at a z-score of 1.5 (κ = 0.69 ± 0.08—concordance in 81 out of 92 reads) and significantly better at a z-score of 2 (κ = 0.84 ± 0.06—concordance in 86 out of 92 reads; p < 0.001).

Figure 2 shows a color-coded overview for both readers, the consensus read at z-score thresholds of 1.5 and 2, and reader confidence values.

Again, the confidence level was significantly higher in temporal lobe epilepsy compared to extratemporal lobe epilepsy for both reviewers at a z-score threshold of 1.5 (reviewer 1 4.2 vs. 3.4, p = 0.002; reviewer 2 4.4 vs. 3.6, p = 0.001) as well as at a z-score threshold of 2 (reviewer 1 4.4 vs. 3.6, p = 0.001; reviewer 2 4.6 vs. 3.7, p < 0.001). Within these subgroups, there was a significantly higher confidence level in temporal lobe epilepsy for reviewer 1 at a z-score threshold of 2 vs. 1.5 (4.4 vs. 4.2, p = 0.02). Confidence levels were also higher in extratemporal lobe epilepsy for both reviewers and in temporal lobe epilepsy for reviewer 2, but these differences failed to reach significance.