Automatic healthy liver segmentation for holmium-166 radioembolization dosimetry

Abstract

Background

For safe and effective holmium-166 (¹⁶⁶Ho) liver radioembolization, dosimetry is crucial and requires accurate healthy liver definition. The current clinical standard relies on manual segmentation and registration of a separately acquired contrast enhanced CT (CECT), a prone-to-error and time-consuming task. An alternative is offered by simultaneous imaging of ¹⁶⁶Ho and technetium-99m stannous–phytate accumulating in healthy liver cells (¹⁶⁶Ho–^99mTc dual-isotope protocol). This study compares healthy liver segmentation performed with an automatic method using ^99mTc images derived from a ¹⁶⁶Ho–^99mTc dual-isotope acquisition to the manual segmentation, focusing on healthy liver dosimetry and corresponding hepatotoxicity. Data from the prospective HEPAR PLuS study were used. Automatic healthy liver segmentation was obtained by thresholding the ^99mTc image (no registration step required). Manual segmentation was performed on CECT and then manually registered to the SPECT/CT and subsequently to the corresponding ¹⁶⁶Ho SPECT to compute absorbed dose in healthy liver.

Results

Thirty-one patients (66 procedures) were assessed. Manual segmentation and registration took a median of 30 min per patient, while automatic segmentation was instantaneous. Mean ± standard deviation of healthy liver absorbed dose was 18 ± 7 Gy and 20 ± 8 Gy for manual and automatic segmentations, respectively. Mean difference ± coefficient of reproducibility between healthy liver absorbed doses using the automatic versus manual segmentation was 2 ± 6 Gy. No correlation was found between mean absorbed dose in the healthy liver and hepatotoxicity.

Conclusions

¹⁶⁶Ho–^99mTc dual-isotope protocol can automatically segment the healthy liver without hampering the ¹⁶⁶Ho dosimetry assessment.

Trial registration: ClinicalTrials.gov, NCT02067988. Registered 20 February 2014. https://clinicaltrials.gov/ct2/show/NCT02067988

Introduction

Radioembolization has been used for many years for treatment of non-operable primary and/or metastatic liver lesions. Among commercially available devices, holmium-166 (¹⁶⁶Ho) microspheres (QuiremSpheres™/QuiremScout™, Quirem Medical BV) allow the use of the same particles for dosimetric assessment during treatment planning and for post-treatment evaluation. In line with conventional radiation oncology, dosimetry plays a key role during these phases.

For hepatic radioembolization, dosimetry is calculated using a partition model which assumes that distinct volumes of interest (VOIs) correspond to different sets of compartments, mainly: healthy liver and tumors [1]. Current clinical practice for ¹⁶⁶Ho radioembolization relies on manual segmentation of these compartments on a previously acquired contrast enhanced CT (CECT) or MRI, and its subsequent registration to the low-dose CT (LDCT) accompanying the SPECT. These two tasks (manual segmentation and registration) present multiple drawbacks, primarily time-consuming and prone to error. Additionally, the manual segmentation and registration introduce an inter-observer variability [2]. These drawbacks might prevent systematic implementation of personalized dosimetry into clinical practice and introduces variability in clinical studies focused on dosimetry. Ultimately this may hinder determination of dose–response relationships and dose–toxicity relationships, and potential treatment outcome prediction.

An alternative to manual segmentation and subsequent manual registration is provided in a protocol based on simultaneous imaging of ¹⁶⁶Ho, simulating the distribution of microspheres [3] in tumors and healthy liver tissue, and technetium-99m (^99mTc) stannous–phytate, accumulating in Kupffer cells located only in healthy tissue (¹⁶⁶Ho–^99mTc dual isotope) [4]. As Kupffer cells are absent in tumorous tissue, this allows a differentiation between tumorous and healthy liver tissue, enabling physiological healthy liver tissue delineation based on [^99mTc]-stannous–phytate uptake [5]. This protocol was applied in a clinical study (HEPAR PLuS) and feasibility of tumor dosimetry on ¹⁶⁶Ho images deriving from this dual-isotope acquisition was previously demonstrated [6].

The aim of this study is to compare dosimetric assessments by manual segmentation (current standard clinical practice) versus the proposed automatically defined healthy liver VOI using the ¹⁶⁶Ho–^99mTc dual-isotope protocol. Furthermore, based on the automatic healthy liver segmentations and the corresponding dosimetry results, a dose–hepatotoxicity relation will be investigated.

Materials and methods

Data population

For all SPECT/CT acquisitions used in this study, informed consent was obtained as part of the HEPAR PLuS study (NCT02067988) [7, 8]. The study has been approved by the institutional review board, and all subjects signed an informed consent form. This study included 31 patients with liver metastases of neuroendocrine tumors (NET), 33 pre-treatment procedures and 38 therapeutic treatments. Baseline characteristics are presented in Table 1.

Table 1 Baseline characteristics of patients with neuroendocrine liver metastases treated in the HEPAR PLuS trial

According to study protocol, for each of the procedures, a SPECT/CT image was acquired after additional activity injection of [^99mTc]-stannous–phytate. For the dataset under consideration, clinical outcomes and lung shunt were previously reported [6, 7].

Images acquisition and reconstruction

All patients were scanned on a Symbia T16 dual-head SPECT/CT scanner (Siemens), using a medium-energy low-penetration collimator, on a 128 × 128 matrix (pixel spacing, 4.8 × 4.8 mm), with 120 angles (15 s per projection) over a non-circular 360° orbit. An energy window centered at 81 keV (15% width), together with an additional energy window centered at 118 keV (12% width) to correct the ¹⁶⁶Ho photopeak data for downscatter using a window-based scatter correction were used [9]. ^99mTc was imaged using a 140 keV (15% width) energy window, with an upper scatter window at 170 keV (12% width) to correct for ¹⁶⁶Ho downscatter. ^99mTc was reconstructed with triple-energy-window scatter correction, using both 118 keV and 170 keV scatter windows. SPECT/CT was acquired following the ¹⁶⁶Ho radioembolization procedure when activity had decayed to approx. 250 MBq, to avoid dead time effects. Five minutes before start of the SPECT scan, ~ 50 MBq of [^99mTc]-stannous–phytate was intravenously injected, to perform the simultaneous ¹⁶⁶Ho–^99mTc dual-isotope acquisition. Reconstruction parameters for both ¹⁶⁶Ho and ^99mTc SPECT were previously analyzed in a phantom study and used in this study [9]. Data were retrospectively assessed to compute the dose in the healthy liver VOI. To this purpose, two approaches were used: manual segmentation and manual rigid registration of CECT on top of SPECT, currently used in clinical practice, and the automatic segmentation on intrinsically registered ^99mTc SPECT using a threshold-based method.

Manual segmentation

For each patient, healthy liver and tumor VOI were manually segmented by an experienced nuclear medicine physician (AJATB) using commercially available software dedicated to ¹⁶⁶Ho radioembolization (Q-Suite™ 2.1, Quirem Medical BV). Delineation of VOI was performed on pre-treatment CECT. Subsequently, the CECT was rigidly registered on the SPECT acquired after the ¹⁶⁶Ho administration procedure through the corresponding LDCT intrinsically registered to the SPECT. Healthy liver VOI was defined by subtracting the manually segmented tumor VOI from total liver volume VOI. Additionally, necrotic tissue was manually segmented as well, to be excluded from healthy liver VOI. Only tumors with a diameter greater than 1 cm were included in the VOI processing. Manual segmentation time for each patient was recorded.

Automatic segmentation

Since [^99mTc]-stannous–phytate accumulates also in the spleen, a preliminary procedure was applied to remove the spleen from the ^99mTc reconstructions. Subsequently, automatic healthy liver segmentation was performed on the ^99mTc reconstruction using a threshold method previously assessed in a phantom study [9]. First, the maximum value in a smoothened version of the ^99mTc image is determined. Then, the original ^99mTc image is thresholded to 40% of this maximum to produce the segmentation. Figure 1 illustrates the manual and automatic segmentation methods within the ¹⁶⁶Ho radioembolization workflow.

Fig. 1

Schematic workflow representing the comparison between manual (left) and automatic (middle and right) segmentation with respect to the dosimetry purpose. The methods have been integrated with corresponding inputs and their timeline within the ¹⁶⁶Ho radioembolization workflow (upper blue arrow)

Dosimetry

To estimate dose in healthy liver VOI, all ¹⁶⁶Ho images were converted to MBq/ml using administered therapeutic activity, on the assumption that all activity was in the SPECT field of view. ¹⁶⁶Ho images were scaled considering only counts within the patient–body mask, automatically delineated on LDCT acquired together with SPECT. Since some patients received a partial liver treatment, ¹⁶⁶Ho pre-treatment images were scaled either using the partial treatment activity or the sum of the activities administered combining the complementary treatments. For all VOIs, dose was computed as:

$${\text{Dose}}\left[ {{\text{Gy}}} \right] = 15.87\left[ {\frac{{{\text{mJ}}}}{{{\text{MBq}}}}} \right]\frac{{ {\text{Activity}}\;{\text{ concentration}}_{{{\text{VOI}}}} \left[ {\text{MBq/ml}} \right]}}{{{\text{VOI}}\;{\text{ density}}\;\left[ {\text{g/ml}} \right]}}$$

where 15.87 mJ/MBq represents the deposited energy due to β decay of 1 MBq ¹⁶⁶Ho. For the liver, a soft tissue density of 1.06 g/cm³ was applied, assuming a homogenous organ density value, constant among patients.

Segmentation evaluation

Healthy liver segmentations obtained using the manual and automatic approach were compared considering the resulting healthy liver volume difference and using two overlapping indices: the Sørensen–Dice coefficient (SDC) and the Hausdorff distance (HD). Since mean absorbed dose within the considered VOI was deemed as a more representative clinical metric, a Bland–Altman plot was used to compare resulting mean absorbed dose computed in the healthy liver manually and automatically segmented. Pearson’s test was used to assess the correlation. As there is no current standard for dose–volume histogram (DVH) reporting, minimum dose to 70% of the volume (D₇₀) and volume receiving at least 50 Gy (V₅₀) were used to compare the two segmentations methods [10]. In each comparative analysis, manual segmentation was considered as reference standard.

Hepatotoxicity was assessed at baseline and up to 12 months, every 3 months, according to the five point standardized scale of hepatotoxicity after radioembolization proposed by Braat et al. [11], and a score of ≥ 3 was considered significant. To assess the dose–hepatotoxicity relation for patients who underwent multiple treatments, mean absorbed doses in the healthy liver resulting from all the treatments received up to the time point under investigation were summed. Dose–hepatotoxicity relation was investigated considering the worst grade hepatotoxicity during the follow-up and cumulative absorbed dose in the healthy liver received by that time point.

To check the statistical difference between resulting doses computed on manual and automatic segmentation (null hypothesis is no difference), a two-sided paired t test (at α = 0.05) was performed.

Results

For 31 subjects and a total of 66 procedures (29 pre- and 37 post-treatment), ¹⁶⁶Ho–^99mTc dual-isotope acquisition was performed and included for segmentation assessment. One patient died within 3 months after ¹⁶⁶Ho radioembolization because of a hypoglycemic crisis caused by an overproducing insulinoma. For another patient who underwent two radioembolization treatments, ¹⁶⁶Ho–^99mTc dual-isotope acquisition was available for only one of the two treatments. These patients were included in the segmentation method comparison, but were excluded in the dose–hepatotoxicity correlation analysis. This resulted in a total of 29 subjects (30 treatments) assessed for hepatotoxicity during follow-up.

Manual segmentation and registration of CECT onto SPECT took a median (IQR) of 30 (15) minutes per patient, while automatic segmentation was instantaneous. Median (IQR) difference between healthy liver volume automatically and manually segmented was 171 mL (341 mL), resulting in a statistically significant difference (p < 0.01). Median (IQR) SDC and HD were 0.8 (0.12) and 8.6 cm (1.4 cm), respectively.

Mean ± standard deviation of health liver absorbed dose was 18 ± 7 Gy and 20 ± 8 Gy for manual and automatic segmentations, respectively. Bland–Altman plot and correlation between dose computed in the healthy liver manually and automatically segmented are shown in Fig. 2. Mean difference ± coefficient of reproducibility between healthy liver absorbed doses using the automatic segmentation versus manual segmentation was 2 ± 6 Gy (Fig. 2A), resulting in limit of agreement of 8 Gy and − 4 Gy. A statistically significant difference in mean healthy liver absorbed doses was found between the two methods (p < 0.01), with the dose computed by automated segmentation, on average, 2 Gy higher. The linear correlation between healthy liver dose computed using these two methods, assessed using Pearson’s correlation coefficient (Fig. 2AB), was 0.92.

Fig. 2

A Bland–Altman plot on the difference between mean absorbed dose. Mean of the difference is depicted by the black solid line, while black dashed lines show limits of agreement. B linear correlation plot between manual and automatic segmentation of the healthy liver VOI with respect to the mean absorbed dose. The solid line depicts linear regression (R² = 0.92), while the dashed lines indicate the ± 95% confidence intervals. Dotted line represents the x = y line. ¹⁶⁶Ho pre-treatment images were scaled considering the therapeutic activity

Mean ± standard deviation of D₇₀ was 10 ± 6 Gy and 11 ± 6 Gy for manual and automatic segmentations, respectively. A Bland–Altman plot and the correlation between D₇₀ are shown in Additional file 1: S1. Mean difference ± coefficient of reproducibility between D₇₀ computed using the automatic segmentation versus manual segmentation was 1 ± 4 Gy (S1A). A statistically significant difference in D₇₀ was found between the two methods, resulting in a mean 1 Gy difference for D₇₀ by automated segmentation (p < 0.01). The linear correlation between D₇₀ computed using these two methods, assessed using Pearson’s correlation coefficient (S1B), was 0.95.

Mean ± standard deviation of V₅₀ was 3 ± 3% and 4 ± 5% for manual and automatic segmentations, respectively. Bland–Altman plot and correlation between V₅₀ computed in the healthy liver manually and automatically segmented are shown in Additional file 1: S2. Mean difference ± coefficient of reproducibility between V₅₀ computed using the automatic segmentation versus manual segmentation was 1 ± 5% (S2A). A statistically significant difference in V₅₀ was found between the two methods (p < 0.01), resulting in a mean 1% difference for V₅₀ by automated segmentation. The linear correlation between V₅₀ computed using these two methods, assessed using Pearson’s correlation coefficient (S2B), was 0.82.

Analysis of hepatotoxicity during the 12 months follow-up after ¹⁶⁶Ho radioembolization, including the number of subjects per each toxicity grade, together with the corresponding mean absorbed dose in the healthy liver manually and automatically segmented is reported in Table 2.

Table 2 Worst hepatotoxicity grade scored according to Braat et al. [11]. Median ± IQR of the absorbed dose in the healthy liver is reported for manual and automatic segmentation

Hepatotoxicity with a significant grade (≥ 3) was found in three patients graded 3, 4 and 5 with a corresponding mean absorbed dose in the manually (and automatically) segmented healthy liver of 19 Gy (20 Gy), 23 (22 Gy) and 30 Gy (32 Gy), respectively. No correlation was found between hepatotoxicity and healthy liver dose. A boxplot representing the healthy liver dose (manually and automatically segmented) per each toxicity grade is depicted in Fig. 3.

Fig. 3

Boxplot representing the healthy liver dose (manually and automatically segmented) for the worst hepatotoxicity grade during follow-up for all 29 patients in the toxicity analysis

Discussion

The ¹⁶⁶Ho–^99mTc dual-isotope protocol offers an automated and reliable alternative to manual segmentation and consequent registration of the healthy liver for future clinical practice. This paves the way to automatic pre-treatment planning and personalized treatment based on healthy liver dosimetry. The ¹⁶⁶Ho–^99mTc dual-isotope protocol can also facilitate rapid post-treatment evaluation, providing insights on dose delivered to the healthy liver and eventually predicting whether significant hepatotoxicity might be encountered during follow-up. Particularly for neuroendocrine tumor patients with long survival times, for whom long-term toxicity is feared [12], healthy liver dose prediction is of even greater importance. The possibility to optimize radioembolization treatment considering the healthy liver dose and corresponding toxicity as limiting factor in order to deliver the maximum tolerable absorbed dose to healthy liver tissue was suggested by Chiesa et al. [13] who investigated this approach for hepatocellular carcinoma patients treated with yttrium-90 glass microspheres.

It is commonly accepted that optimal radioembolization requires accurate dosimetry of both tumor and healthy liver [14, 15]. The mean absorbed dose difference between automatically and manually segmented healthy liver is limited to 2 Gy, which, even though statistically significant, was considered acceptable for healthy liver assessment by experienced nuclear medicine physicians. The same considerations apply to the D₇₀ and V₅₀ used to assess the DVH. These results are in line with the results from tumor dosimetry with the same ¹⁶⁶Ho–^99mTc dual-isotope protocol, also 2 Gy overestimation [6]. With these limited differences in mind, ¹⁶⁶Ho–^99mTc dual-isotope protocol seems a viable imaging protocol to enhanced implementation of automated personalized dosimetry in daily clinical practice. However, at this time, the possibility to perform tumor dose assessment is subject to manual segmentation and registration of lesions by physicians. Thus, as for healthy liver, dosimetry would benefit from automatic tumor segmentation. Despite ¹⁶⁶Ho uptake occurring primarily in tumors (for hypervascular lesions), a minor amount of ¹⁶⁶Ho is distributed in healthy liver tissue. For this reason, a robust thresholding approach to segment tumors based on the ¹⁶⁶Ho image alone is challenging to implement. The use of the ^99mTc image derived from the ¹⁶⁶Ho–^99mTc dual-isotope protocol for automatic tumors segmentation would rely on automatic identification of cold spots (considered as tumors due to the lack of Kupffer cells present in parenchyma). This approach presents several challenges since the limited SPECT resolution would play a bigger role in identification of small volumes of interest. The resolution will limit detection of small tumors, while miss identification of cold spots as tumors could potentially be benign lesions (e.g., liver cysts). Additionally, to perform automatic tumor dosimetry, the perfused volume during treatment should be identified. While the ¹⁶⁶Ho image itself provides an indication of the perfused volume, this task is not straightforward (especially for lobar treatments) and brings a further challenge to automatic tumor dosimetry using the ¹⁶⁶Ho–^99mTc dual-isotope protocol.

Besides saving time (~ 30 min per patient when using the manual segmentation versus instantaneous with this ¹⁶⁶Ho–^99mTc dual-isotope protocol), another straightforward benefit of the ¹⁶⁶Ho–^99mTc dual-isotope protocol is the potential to avoid of inter-observer variability when conducting dosimetry [2]. Several large studies have shown the importance of dosimetry in patient outcome [16, 17], and clinical studies currently enrolling participants implement dosimetric endpoints. Thus, limiting underlying errors (e.g., registration errors and inter-observer variability) will become more important for data generalizability in the future. In this study, some of these variations become apparent, as despite of overlapping indices like SDC and HD, which are easy to implement and useful summary measures of spatial overlap, they poorly outline the considerations assessing the segmentation methods (manual versus automatic) for the purpose of ¹⁶⁶Ho radioembolization dosimetry. An explanatory example showing how overlapping indices might lead to partial conclusions on the resulting dosimetry is depicted in Fig. 4.

Fig. 4

Examples of automatic segmentation outcome. A–D. A 56-year-old male subject treated with 3850 MBq of ¹⁶⁶Ho microspheres in the right liver. E–H. A 65-year-old male subject treated with 4982 MBq of ¹⁶⁶Ho microspheres in the right liver. Panels A and E, and B and F, show the axial view of ¹⁶⁶Ho and ^99mTc SPECT/CT, respectively, while panels C and G show the coronal view of the LDCT acquired with the SPECT. Manual (solid green) and automatic (dashed magenta) segmentation is depicted. Upper panels (A–C, Sørensen–Dice coefficient = SDC = 0.62), show a clear discrepancy between segmentations, which is due to non-rigid deformation of liver because of breathing. Segmentations are quite comparable in the bottom panels (E–G, SDC = 0.84). Panels D and H show the resulting cumulative dose–volume histogram (DVH). Despite the segmentation differences, the corresponding dose difference is below 2 Gy

In this study, all patients presented a mean healthy liver absorbed dose (manually or automatically segmented) < 35 Gy (range 3–32 Gy), below the reported acceptable healthy liver dose limits reported for patients with different tumor types treated with radioembolization [13, 18, 19]. However, due to the limited REILD occurrence within this study (1 case) and low incidence of significant hepatotoxicity (1 case), no correlation was found and no strong conclusion can be drawn on maximum tolerable healthy liver dose. This is in line with previous results by Ebbers et al. [20] in which no significant relationship between absorbed dose in treated healthy liver and biochemical toxicity was found when assessing a NET patient population treated with yttrium-90 glass microspheres. To date, no clear dose–toxicity relationships have been described in radioembolization in general; advised thresholds for healthy liver dose are based on limited evidence and often lack validation [21].

Several limitations apply to this study. Primarily, the healthy liver segmentation manually performed on baseline anatomical image and rigidly registered on SPECT is known to be far from ideal (as shown in Fig. 4A–C); even though it is current clinical practice, the reference standard (‘ground truth’) to which the automatic segmentation outcome was compared is limited. Also, the automatic segmentation method, which relies on a fixed threshold for all patients, can lead to suboptimal segmentation for some patients. [^99mTc]-stannous–phytate accumulates in Kupffer cells only, representing healthy liver; however, its uptake is not homogenous and does not provide any information regarding regional liver function. Furthermore, an intrinsic limitation of ¹⁶⁶Ho–^99mTc dual-isotope approach is the limited resolution of SPECT, which does not allow for the identification of small lesions, widely present in NET patients. Additionally, the definition of a clear dose–hepatotoxicity relation is hampered by the limited population size (and number of events) and hepatotoxicity following radioembolization is related to multiple risk factors such as previous liver directed treatments or intra-arterial therapies, tumor burden and toxicity of concomitant medications [11].

¹⁶⁶Ho–^99mTc dual-isotope protocol is a promising and rapid protocol for automated imaging segmentation; however, several challenges still need to be addressed, especially concerning automated tumor dosimetry. A clinical validation study of the ¹⁶⁶Ho–^99mTc dual-isotope protocol, including more patients, should be performed to confirm the presented promising data and prove its ease in the clinical workflow.

Conclusions

In conclusion, ¹⁶⁶Ho–^99mTc dual-isotope imaging allows automatic segmentation of the healthy liver tissue using a thresholding method without compromising assessment of healthy liver absorbed dose. ¹⁶⁶Ho–^99mTc dual-isotope imaging paves the way for automated partition model-based activity calculation for ¹⁶⁶Ho radioembolization, feasible in clinical practice. Prospective validation data are needed.