A novel statistical analysis method to improve the detection of hepatic foci of 111In-octreotide in SPECT/CT imaging
© Magnander et al. 2016
Received: 12 November 2015
Accepted: 12 January 2016
Published: 19 January 2016
Low uptake ratios, high noise, poor resolution, and low contrast all combine to make the detection of neuroendocrine liver tumours by 111In-octreotide single photon emission tomography (SPECT) imaging a challenge. The aim of this study was to develop a segmentation analysis method that could improve the accuracy of hepatic neuroendocrine tumour detection.
Our novel segmentation was benchmarked by a retrospective analysis of patients categorized as either 111In-octreotide positive (111In-octreotide(+)) or 111In-octreotide negative (111In-octreotide(−)) for liver tumours. Following a 3-year follow-up period, involving multiple imaging modalities, we further segregated 111In-octreotide-negative patients into two groups: one with no confirmed liver tumours (111In-octreotide(−)/radtech(−)) and the other, now diagnosed with liver tumours (111In-octreotide(−)/radtech(+)). We retrospectively applied our segmentation analysis to see if it could have detected these previously missed tumours using 111In-octreotide. Our methodology subdivided the liver and determined normalized numbers of uptake foci (nNUF), at various threshold values, using a connected-component labelling algorithm. Plots of nNUF against the threshold index (ThI) were generated. ThI was defined as follows: ThI = (c max − c thr)/c max, where c max is the maximal threshold value for obtaining at least one, two voxel sized, uptake focus; c thr is the voxel threshold value. The maximal divergence between the nNUF values for 111In-octreotide(−)/radtech(−), and 111In-octreotide(+) livers, was used as the optimal nNUF value for tumour detection. We also corrected for any influence of the mean activity concentration on ThI. The nNUF versus ThI method (nNUFTI) was then used to reanalyze the 111In-octreotide(−)/radtech(−) and 111In-octreotide(−)/radtech(+) groups.
Of a total of 53 111In-octreotide(−) patients, 40 were categorized as 111In-octreotide(−)/radtech(−) and 13 as 111In-octreotide(−)/radtech(+) group. Optimal separation of the nNUF values for 111In-octreotide(−)/radtech(−) and 111In-octreotide(+) groups was defined at the nNUF value of 0.25, to the right of the bell shaped nNUFTI curve. ThIs at this nNUF value were dependent on the mean activity concentration and therefore normalized to generate nThI; a significant difference in nThI values was found between the 111In-octreotide(−)/radtech(−) and the 111In-octreotide(−)/radtech(+) groups (P < 0.01). As a result, four of the 13 111In-octreotide(−)/radtech(+) livers were redesigned as 111In-octreotide(+).
The nNUFTI method has the potential to improve the diagnosis of liver tumours using 111In-octreotide.
KeywordsSPECT/CT 111In-octreotide Diagnosis Liver tumours Neuroendocrine
Timely and precise detection of metastatic colonies in the liver of a cancer patient is crucial if we are to make the best treatment choices . Metastatic disease may indicate that a different curative regimen, or perhaps a palliative option, is now the best course of action [2–5]. Unfortunately, our ability to make these decisions is hampered by the difficulty that we face in detecting small masses, with low contrast and noise .
Single photon emission tomography (SPECT), using a radiolabelled mimic of somatostatin, 111In-octreotide, has become the established methodology with which to image somatostatin receptor (sstr)-positive tumours [7, 8] in patients with neuroendocrine tumours. Together with morphological imaging techniques, such as computed tomography (CT) and magnetic resonance imaging (MRI), it has become an important tool for tumour visualization, staging, and evaluation of somatostatin receptor status . In addition, one of the major eligibility criteria for patients undergoing therapy with either 177Lu- or 90Y-labelled somatostatin analogues is that their tumour should bind more 111In-octreotide than normal liver [10–12]. The purpose of this study was to develop a complementary method that could help the physician in reaching their diagnosis of sstr-tumour involvement using SPECT images of the liver.
Despite significant observer variability, nuclear medicine still, to a large extent, depends on the subjective decision-making of a single physician [13, 14]. Unfortunately, quantitative analyses that are simple to use, and easy to understand for the observer, are rarely implemented in this field. Nevertheless, recent success with computer-assisted diagnosis in the detection of bone metastases [15, 16] may prompt a rethink for neuroendocrine tumours. A key aspect of these systems is their necessity for a ‘library’ of patient data; in essence, the algorithm needs to be taught how to recognize the tumour. With this input, it should be possible to develop more straightforward, and objective methodologies, with which to guide decision-making.
Due to the limited spatial resolution of the gamma camera, small tumours, with only a modest uptake of the radiolabel, will fall beneath the limit of detection. A second problem that limits resolution is noise in the system. Voxel clustering may give a false impression of a mass that has bound the radiolabel, i.e. a false positive signal is generated. The inherent loss in specificity that this generates further diminishes the chances of authentically locating a metastatic focus. Our experimental methodology is based on a statistical approach with the assumption that the distributions of uptake foci differ between healthy livers and livers with tumour involvements. Starting with the maximum voxel value within the liver, and successively calculating the number of segmented uptake foci at decreasing threshold values, a graph of the number of uptake foci, versus threshold value, is obtained. We will, in this study, demonstrate that the number of disjointed segmented uptake foci, as a function of decreasing threshold value for normal liver (with no true uptake foci), can be displayed as a symmetrical bell-shaped curve. When authentic foci are present, this curve is shifted and/or compressed towards decreased threshold values. This shift can be exploited to discriminate between authentic and artifactual foci. We then apply this method in a retrospective analysis of 53 randomly selected patients, previously diagnosed by 111In-octreotide SPECT to be clear of liver metastases. Within this group are a number of patients that, during a 3-year follow-up, were confirmed by other methods to display liver metastases. For these patients, we used our experimental methodology to address whether it would have helped the clinician to reach a positive 111In-octreotide diagnosis for liver involvement, rather than the negative diagnoses that were reported.
Segmentation of uptake foci
Automatic segmentation of uptake foci in liver tissue at different voxel threshold values is a methodology developed for the analysis of SPECT/CT data using 111In-octreotide. Raw image data (128 × 128, 120 projections) were reconstructed according to the standard clinical protocol, i.e. using ordered subset expectation maximum (OSEM) reconstruction, with two iterations and ten subsets. In a departure to this protocol, the resulting volume was then unfiltered; the justification for this was to improve our chances of finding small lesions. The liver volume of interest (VOI) was segmented from SPECT, but in some SPECT investigation, the high uptake in surrounding tissues hampered the automatic segmentation. In these situations, the segmentations were performed in the CTs. The segmentations performed from the SPECT, or CT, used either an isosurface, a region growing, or a graphics processing unit (GPU) accelerated level set algorithm. Manual editing was used to refine the segmentation. The liver VOI was then thresholded at values between 0 and the maximum voxel value. At each threshold value, the number of uptake foci (NUF), i.e. the number of connected regions unlinked to other regions, was determined using a technique known as connective-component labelling . Connective-component labelling is an algorithm where subsets of connected components (here, regions of connected voxels) are uniquely labelled, enabling calculation of the number of subsets, i.e. the NUF in our method.
The normalized number of uptake foci versus threshold index (nNUFTI) method
where C max is the maximal voxel value in the VOI and C thr is the voxel threshold value. The NUF can therefore be described as a function of ThI, ranging from 0 to 1. By normalizing the NUF (nNUF) to the maximal NUF in a liver VOI, it is possible to display and then compare individual 111In-octreotide SPECT data (by comparing nNUF versus ThI (nNUFTI)).
Creation of a visualization tool for the observer
A parallel aim was to integrate the nNUFTI method with a visualization tool for the observer. This was performed by incorporating a slider along the ThI-axis, which provides a real-time visual representation of the corresponding uptake foci in the 3D liver VOI. Unconnected uptake foci were assigned different colours, from a palette of 256. This enabled the observer to pinpoint suspicious lesions by closely studying their noise structure and the clustering of uptake foci. Additionally, the displayed 3D volume of disconnected uptake foci can be analysed contemporaneously with CT and MRI data. The nNUFTI programme was written in c++ and implemented into PhONSAi (the medical Physics, Oncology and Nuclear medicine image research platform at Sahlgrenska Academy). Using the PhONSAi platform, the segmentation of the liver, together with CT data processing, was performed in parallel on the graphic card (CUDA) to obtain a rapid segmentation performance. An experienced clinical observer oversaw the graphical layout of the nNUFTI programme.
Determination of optimal value for the normalized number of uptake foci
To determine the optimal value of the nNUF that allows the greatest separation of ThI between the 111In-octreotide(−)/radtech(−), and 111In-octreotide(+) groups, the t score in Student’s t test was used. The nNUF with the highest t score was used to determine the ThI values in further 111In-octreotide analyses using the nNUFTI method.
Analysis of the 111In-octreotide-negative patients with the nNUFTI method
Using the nNUFTI method, the ThI was determined for all patients using the value of nNUF as previously described. For the 111In-octreotide(−)/radtech(−) group, it was confirmed that ThI was dependent on the mean activity concentration; therefore, ThI was normalized (nThI) by the function that best describes this dependency. Comparison of the nThI for the 111In-octreotide(−)/radtech(−) and 111In-octreotide(−)/radtech(+) groups was conducted by Student’s t test. P < 0.05 was considered significant. Lastly, the positive predictive value (PPV) for the individual nTHI values in the 111In-octreotide(−)/radtech(+) group was calculated.
Of the first 80 selected 111In-octreotide(−) patients, 27 had to be excluded due to incomplete follow-up data. Forty of the remaining 53 patients were diagnosed as having no sign of liver tumours within their follow-up time. These were designated the 111In-octreotide(−)/radtech(−) group. The remaining 13 patients had their liver tumours detected by either CT or ultra sound, designated as the 111In-octreotide(−)/radtech(+) group.
Characterization of the octreotide(−)/radtech(+) livers, including results from the nNUFTI analyses
Time of diagnosis (years)
Number of metastases
NET, origin unknown
Medullary thyroid cancer
Medullary thyroid cancer
Midgut carcinoid, cervix cancer
We now report a novel and improved segmentation method, with a proven ability to produce real-time images of uptake foci, for immediate evaluation by the observer. The nNUFTI display provides the observer with a quality-control parameter for tumour involvement, with an additional probability value provided by the nThI.
The nNUFTI algorithm that we created is based on the assumption that 111In-octreotide-positive tumours will skew the nNUFTI curves towards higher ThI values. Our results also showed that both large 111In-octreotide-positive tumour burdens, as well as modest tumour involvement, could be easily detected by their compression of the nNUFTI curve. These data emphasize the value of this methodology in the early detection of metastatic liver disease. To evaluate if this methodology could outperform conventional 111In-octreotide imaging, we randomly selected 53 111In-octreotide negative patients and provided follow-up for the detection of emergent, late, liver tumours. Thirteen of these patients were found to display tumours in CT, MRI, PET/CT, or ultra sound investigations, i.e. 25 % of the 111In-octreotide-negative patients were false negatives. The nNUFTI method was able to yield a statistically significant separation between the two groups, suggesting that the nNUFTI algorithm can, without bias, detect emergent liver tumours.
The main obstacles to tumour detection using SPECT imaging include difficulty in discriminating between uptake of the radiolabel by normal tissue versus the tumour due to a low tumour-to-normal activity concentration ratio (TNC), poor resolution, and noise. These confounding factors are the same, whether using conventional observation or the nNUFTI method. To mitigate these issues, we defined a quantitative measure for tumour detection by comparing 40 healthy livers with 10 111In-octreotide-positive livers. Five of the tumour-positive livers were selected because they carried a modest tumour burden; these are the patients whose cancers are ordinarily most difficult to detect. We determined that a nNUF value of 0.25, at the right side of the nNUFTI curve, provided the greatest discrimination between groups. Our results also highlighted the fact that 111In-octreotide is unevenly distributed through the liver (Fig. 2), with the highest uptake displayed initially in the segment representing the right side of the lobe; this non-uniform distribution of activity was observed for most livers. When the TNC is constant, as compared with a regional liver activity distribution, tumours localized to a segment of high-activity concentration will compress the entire nNUFTI curve, whereas a tumour localized to a low activity concentration segment might only compress the nNUFTI curve on the right side. This non-uniformity of hepatic activity concentration increases the sensitivity of the nNUFTI method on the right side of the curve.
The influence of noise on the nNUFTI method was demonstrated by the strong nThI versus voxel signal correlation. This was corrected for, using a polynomial function. While different models can be fitted to the data, our choice was based on the best-fit approach. Future modelling studies should be performed to codify this method.
We noted a variation in the mean voxel signal for 111In-octreotide between livers due to different radiolabel pharmacokinetics, non-standard sensitivity of the gamma cameras, and different SPECT measurement times. In particular, we noted a marked difference for ThI when using our two gamma cameras; this introduced a 25 % difference in sensitivity for 111In. A depressed sensitivity increased the ThI, which then had to be corrected using the adjusted nThI. This correction eliminated the variability of the camera and the duration of measurement.
The probability estimator that we derived for tumour involvement is intended to guide the observer in their diagnostic decision-making. SPECT data with nThIs of between 0.64 and 0.69 appeared to be the most problematic in terms of obtaining an accurate diagnosis. For these patients, it is recommended that follow-up includes other tests, i.e. CT, MRI, ultra sound, PET/CT, or additional SPECT/CT. For lower nThIs, it seems most probable that additional 111In-octreotide investigations will be of less value, either because there are no tumours to be found or, if present, they are 111In-octreotide negative. A higher nThIs (>0.70) is a strong indicator of tumour burden (Fig. 6), although it may be difficult to visualize the tumours even with the developed visualization tool. Comparing Figs. 3 and 7, the tumours marked in the octreotide(−) liver in Fig. 7 are hardly detectable even with knowledge of their localization. Nevertheless, the nNUFTI method reported a value of nThI as high as 0.74, well above any radtech(−) liver and showing the sensitivity of the nNUFTI method. In contrast, the tumour in the octreotide(+) liver in Fig. 3, with an nThI of 0.84, is well visualized.
The present study followed the standard reconstruction protocols used in the clinic, i.e. OSEM with two iterations and ten subsets; ultimately, this may not be the optimal setting for the nNUFTI method. Further modelling studies using simulated tumours in the non-malignant livers, and different reconstruction settings, might be of value in determining the best reconstruction method to use with nNUFTI. Necessarily, for obtaining realistic noise, the modelling studies should be performed by Monte Carlo simulations of the tumours into the sinograms and thereafter reconstruction.
The tumours that could be detected by the nNUFTI method were also diagnosed contemporaneously with CT or ultrasound (US). Those tumours not detected by nNUFTI were only found by CT or US, 1 to 3 years after the initial 111In-octreotide diagnosis. These late arising tumours were not detected by a later 111In-octretide investigation. These data would suggest that the expression of the somatostatin receptor might be low for these tumours. Two of these tumours were medullary thyroid cancers, and one was a lung carcinoid, both known to have a lower uptake of 111In-octreotide , i.e. they were poor candidates for detection by 111In-octreotide.
Early detection of liver tumours is of huge benefit to the patient in terms of managing their treatment options. Current treatment for neuroendocrine tumours comprises different regimens. One used alternative, that has achieved some attention, is treatment with the radionuclide-labelled somatostatin analogues, 177Lu-DOTATATE and 90Y-DOTATOC [10, 19, 20]. One criterion for selecting patients for these treatments is that the uptake of the radionuclide is higher in the tumours than in normal liver tissue. However, this estimation is difficult to ascertain for small tumours, and therefore, these treatments tend to be biased towards patients with more advanced disease. However, it would be preferable to use these radionuclides to treat the disease at an earlier stage . Future studies will also be performed to analyse whether this method can be used to follow treatment response in all NUF regions, thereby estimating the metastatic cure probabilities for liver tumours [21, 22].
While the present work focused on SPECT data of 111In-labelled somatostatin analogues in the liver, the method that we developed might be applicable to other volumes of interest in the patient and is also practicable for all diagnostic radionuclides, and in extension, to PET-imaging.
We verified the utility of a novel method (nNUFTI) with which to convincingly detect observer defined 111In-octreotide-positive tumours, as well as non-visualized tumours, in SPECT images. Our data indicates that the nNUFTI algorithm has the potential to become a useful analytical tool with which to complement, and improve, the conventional diagnosis of liver tumours using 111In-octreotide.
Compliance with ethical standards
This work was funded by the Swedish Cancer Society, the Swedish Radiation Safety Authority, the King Gustav V Jubilee Clinic Cancer Research Foundation, and the Swedish Federal Government under ALF agreement. This retrospective study was approved by the Regional Ethical Review Board in Gothenburg and performed in accordance with the Declaration of Helsinki and national regulations. The need for written informed consent was waived.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Elias D, Lefevre JH, Duvillard P, Goere D, Dromain C, Dumont F, et al. Hepatic metastases from neuroendocrine tumors with a “thin slice” pathological examination: they are many more than you think. Ann Surg. 2010;251:307–10.PubMedView ArticleGoogle Scholar
- Wong KK, Cahill JM, Frey KA, Avram AM. Incremental value of 111-in-pentetreotide SPECT/CT fusion imaging of neuroendocrine tumors. Acad Radiol. 2010;17(3):291–7.PubMedView ArticleGoogle Scholar
- Castaldi P, Rufini V, Treglia G, Bruno I, Perotti G, Stifano G, et al. Impact of 111In-DTPA-octreotide SPECT/CT fusion images in the management of neuroendocrine tumours. Radiol Med. 2008;113(7):1056–67.PubMedView ArticleGoogle Scholar
- Krausz Y, Keidar Z, Kogan I, Even-Sapir E, Bar-Shalom R, Engel A, et al. SPECT/CT hybrid imaging with 111In-pentetreotide in assessment of neuroendocrine tumours. Clin Endocrinol. 2003;59(5):565–73.View ArticleGoogle Scholar
- Hillel PG, van Beek EJ, Taylor C, Lorenz E, Bax ND, Prakash V, et al. The clinical impact of a combined gamma camera/CT imaging system on somatostatin receptor imaging of neuroendocrine tumours. Clin Radiol. 2006;61(7):579–87.PubMedView ArticleGoogle Scholar
- Rahmim A, Zaidi H. PET versus SPECT: strengths, limitations and challenges. Nucl Med Commun. 2008;29(3):193–207.PubMedView ArticleGoogle Scholar
- Krenning EP, Kwekkeboom DJ, Bakker WH, Breeman WA, Kooij PP, Oei HY, et al. Somatostatin receptor scintigraphy with [111In-DTPA-D-Phe1]- and [123I-Tyr3]-octreotide: the Rotterdam experience with more than 1000 patients. Eur J Nucl Med. 1993;20(8):716–31.PubMedView ArticleGoogle Scholar
- Lu SJ, Gnanasegaran G, Buscombe J, Navalkissoor S. Single photon emission computed tomography/computed tomography in the evaluation of neuroendocrine tumours: a review of the literature. Nucl Med Commun. 2013;34(2):98–107.PubMedView ArticleGoogle Scholar
- van Essen M, Sundin A, Krenning EP, Kwekkeboom DJ. Neuroendocrine tumours: the role of imaging for diagnosis and therapy. Nat Rev Endocrinol. 2014;10(2):102–14.PubMedView ArticleGoogle Scholar
- Kwekkeboom DJ, de Herder WW, Kam BL, van Eijck CH, van Essen M, Kooij PP, et al. Treatment with the radiolabeled somatostatin analog [177 Lu-DOTA 0, Tyr3]octreotate: toxicity, efficacy, and survival. J Clin Oncol. 2008;26(13):2124–30.PubMedView ArticleGoogle Scholar
- Balon HR, Brown TL, Goldsmith SJ, Silberstein EB, Krenning EP, Lang O, et al. The SNM practice guideline for somatostatin receptor scintigraphy 2.0. J Nucl Med Technol. 2011;39(4):317–24.PubMedView ArticleGoogle Scholar
- Bombardieri E, Ambrosini V, Aktolun C, Baum RP, Bishof-Delaloye A, Del Vecchio S, et al. 111In-pentetreotide scintigraphy: procedure guidelines for tumour imaging. Eur J Nucl Med Mol Imaging. 2010;37(7):1441–8.PubMedView ArticleGoogle Scholar
- Sadik M, Suurkula M, Höglund P, Järund A, Edenbrandt L. Quality of planar whole-body bone scan interpretations—a nationwide survey. Eur J Nucl Med Mol Imaging. 2008;35(8):1464–72.PubMedView ArticleGoogle Scholar
- Koopmans KP, Neels ON, Kema IP, Elsinga PH, Links TP, de Vries EG, et al. Molecular imaging in neuroendocrine tumors: molecular uptake mechanisms and clinical results. Crit Rev Oncol Hematol. 2009;71(3):199–213.PubMedView ArticleGoogle Scholar
- Sadik M, Hamadeh I, Nordblom P, Suurkula M, Höglund P, Ohlsson M, et al. Computer-assisted interpretation of planar whole-body bone scans. J Nucl Med. 2008;49(12):1958–65.PubMedView ArticleGoogle Scholar
- Ulmert D, Kaboteh R, Fox JJ, Savage C, Evans MJ, Lilja H, et al. A novel automated platform for quantifying the extent of skeletal tumour involvement in prostate cancer patients using the bone scan index. Eur Urol. 2012;62(1):78–84.PubMedPubMed CentralView ArticleGoogle Scholar
- Chang WY, Chiu CC, Yang JH. Block-based connected-component labeling algorithm using binary decision trees. Sensors (Basel). 2015;15(9):23763–87.View ArticleGoogle Scholar
- Forssell-Aronsson E, Bernhardt P, Nilsson O, Tisell LE, Wangberg B, Ahlman H. Biodistribution data from 100 patients i.v. injected with 111In-DTPA-D-Phe1-octreotide. Acta Oncol. 2004;43(5):436–42.PubMedView ArticleGoogle Scholar
- Bodei L, Cremonesi M, Grana CM, Fazio N, Iodice S, Baio SM, et al. Peptide receptor radionuclide therapy with (1)(7)(7)Lu-DOTATATE: the IEO phase I-II study. Eur J Nucl Med Mol Imaging. 2011;38(12):2125–35.PubMedView ArticleGoogle Scholar
- Svensson J, Berg G, Wängberg B, Larsson M, Forssell-Aronsson E, Bernhardt P. Renal function affects absorbed dose to the kidneys and haematological toxicity during (177)Lu DOTATATE treatment. Eur J Nucl Med Mol Imaging. 2015;42(6):947–55.PubMedPubMed CentralView ArticleGoogle Scholar
- Bernhardt P, Ahlman H, Forssell-Aronsson E. Model of metastatic growth valuable for radionuclide therapy. Med Phys. 2003;30(12):3227–32.PubMedView ArticleGoogle Scholar
- Bernhardt P, Ahlman H, Forssell-Aronsson E. Modelling of metastatic cure after radionuclide therapy: influence of tumor distribution, cross-irradiation, and variable activity concentration. Med Phys. 2004;31(9):2628–35.PubMedView ArticleGoogle Scholar