EMIM 2019
To search for a specific ID please enter the hash sign followed by the ID number (e.g. #123).

Data Processing & Quantification

Session chair: Adriana Tavares (Edinburgh, UK); Athanasios Zacharopolous (Heraklion, Greece)
 
Shortcut: PW10
Date: Thursday, 21 March, 2019, 12:45 p.m.
Room: ALSH | level 0,BOISDALE | level 0,CARRON | level +1,DOCHART | level +1
Session type: Poster

Contents

Click on an contribution to preview the abstract content.

201

Fully Automated Radiomic Analysis for Tumor Detection and Classification in Ultrasound Images (#415)

Zuzanna A. Magnuska1, Tatjana Opacic1, Severine Iborra2, Elmar Stickeler2, Fabian Kiessling1, 3, Benjamin Theek1, 3

1 University Clinic Aachen, RWTH Aachen University, Institute for Experimental Molecular Imaging, Aachen, North Rhine-Westphalia, Germany
2 University Clinic Aachen, RWTH Aachen University, Department of Obstetrics and Gynecology, Aachen, North Rhine-Westphalia, Germany
3 Fraunhofer MEVIS, Bremen, Bremen, Germany

Introduction

The goal of radiomics is to extract quantitative features from radiological images and correlate them with clinical findings. This might improve precision medicine by developing algorithms supporting physicians in diagnosis and therapy. Thus far, only a few radiomic studies on ultrasound (US) images have been published [1,2]. Therefore, we created an user-independent workflow to conduct a radiomic analysis of US images and assessed its capability to automatically differentiate xenograft tumor models. Moreover, the developed algorithms were adapted to analyze clinical breast cancer US images.

Methods

US B-mode images of 3 different xenograft tumor models (lung cancer (n=3), ovarian cancer (n=3), SCC (n=3)) and clinical breast cancer types (fibroadenoma (n=6), HER2-positive (n=2), TNB (n=7)) were retrospectively analyzed. Cascade classifiers (CCs) were trained for an automated tumor detection (Fig. 1A) [3]. The tumor segmentation was based on an active contour model and morphological operations. From the segmented region, a total of 230 intensity-based, textural and wavelet-based features were mined (Fig. 1C). Feature stability, discriminative power and user-independence were evaluated to identify 3 non-correlating traits, composing the radiomic signature (RS). Using the RS as input, a k-NN based classification algorithm was evaluated with the leaving-one-out-cross-validation scheme.

Results/Discussion

The developed algorithm for the automated detection and segmentation of tumors achieved a high detection accuracy (89% of correct tumor detections), and automated segmentations overlapped with the manual segmentations (81% ± 24% overlap) (Fig. 1B). The developed imaging biomarker extraction and selection algorithm identified the following 3 non-correlating features as the RS: median (intensity-based feature), correlation (textural feature) and short run emphasis (wavelet-based feature). The tumor classification model assigned 80% of the analyzed images to the correct tumor type (p=0.8 [95% CI 0.6-0.9]) (Fig. 1C). The application of the established radiomic analysis on breast cancer US images resulted, thus far, in a similar detection accuracy (87% of correct tumor detections) (Fig. 2).

Conclusions

Our results show that a radiomic analysis of US images can be performed to classify tumors. Initial results on clinical data suggest that the established concept can be translated to the clinical setting as well. The developed automated workflow may assist clinicians in a more robust and reliable tumor recognition, characterisation and differentiation. More extensive evaluation of clinical data will be performed in future studies.

References

[1] Lee, S.E., et al., Scientific Reports 2018. 8(1): p. 13546.

[2] Guo Y., et al. Clin Breast Cancer, 2017.

[3] Viola P. and Jones M. CVPR, 2001.

Acknowledgement

This work has been supported by the German Research Foundation (KI1072/11-1).

 

The retrospective study of clinical breast cancer US images was approved by the ethics comity of RWTH Aachen University, Germany.

Fully Automated Radiomic Analysis of Experimental Tumor Models
Automated Tumor Detection in Clinical Breast Cancer US Images
Keywords: biomedical image processing, data analysis, texture analysis, automated segmentation, radiomics
202

Developing a customizable workflow engine for storing, sharing, processing and reusing medical images for preclinical imaging facilities (#140)

Sara Zullino1, Alessandro Paglialonga1, Walter Dastrù1, Alessandra Viale1, Silvio Aime1, Dario L. Longo1

1 University of Torino, Molecular Imaging Center, Department of Molecular Biotechnology and Health Sciences, Torino, Italy

Introduction

Medical imaging data is a highly valuable resource for research on diagnostics, epidemiology and drug development. However, a common limitation relies on the lack of common tools for sharing, processing and reusing medical imaging data acquired with different instrumentation among several imaging centers. Our aim is to overcome this urgent need through the integration of an open-source archiving platform with customizable tools for automated image processing. This workflow will allow the preclinical research community to store, process and reuse a large number of medical imaging data.

Methods

We have developed Python/Matlab-based tools for exporting, processing and archiving preclinical images exploiting the built-in Pydicom library [1]. These tools interface with the remotely accessible database XNAT [2], a widely used open source platform for managing, sharing and processing medical imaging data, via XNAT Python clients, xnatpy and Pyxnat [3, 4]. Since preclinical instrumentation adopts a proprietary format for their data, these tools can convert raw images to DICOM format from different vendors (MRI Bruker and Aspect Imaging) or can directly import DICOM images for several imaging modalities (PET/SPECT/CT/OI/OA). Dedicated tools have been implemented for automatically storing whole experiment datasets including annotations for experimental groups and timepoints.

Results/Discussion

The workflow is based on the following steps (Fig.1):

  1. a retrieving/archiving step to collect multiple imaging datasets acquired through several modalities;
  2. a Bruker/Aspect to DICOM format converter to import images to XNAT;
  3. an image processing step accepting as input either DICOM (3a) or raw data (3b) to produce parametric images related to biological aspects.

The workflow can decipher preclinical MRI raw data and convert proprietary images to DICOM, storing the parameters in both standard and private DICOM tags. To match our data organization, a XNAT feature to extend an existing data type using custom variables has been used. XNAT pipelines have been implemented to import large datasets arising from several studies/patients/sessions and to process MR images. XNAT can support the storage of several files at the resource level, such as binary masks, parametric images, report files and annotated information (Fig. 2). 

Conclusions

A user-friendly, customizable workflow has been developed to store and process image datasets from preclinical imaging centers using a XNAT-based archive system. This approach complies with the FAIR (Findable, Accessible, Interoperable and Reusable) guidelines. This engine will be available for the preclinical research community, allowing to access and share image data among preclinical imaging centers.         

References

[1]         D. Mason, “Pydicom: An Open Source DICOM Library,” in Medical Physics, 2011, vol. 38, no. 6, p. 3493.

[2]         D. S. Marcus, T. R. Olsen, M. Ramaratnam, and R. L. Buckner, “The extensible neuroimaging archive toolkit: An informatics platform for managing, exploring, and sharing neuroimaging data,” Neuroinformatics, vol. 5, no. 1, pp. 11–33, 2007.

[3]         Hakim Achterberg, “XNAT Python Client,” https://xnat.readthedocs.io/en/latest/

[4]         Y. Schwartz et al., “PyXNAT: XNAT in Python,” Front. Neuroinform., vol. 6, p. 12, May 2012.

Acknowledgement

European Union’s Horizon 2020 research and innovation programme under grant agreement No 654248 (CORBEL project), grant No 667510 (GLINT project) and from Compagnia San Paolo project (Regione Piemonte, grant #CSTO165925).

Figure 1
Schematic workflow for image archiving and processing.
Figure 2
Snapshots of the image session webpage on the XNAT platform: a) custom variables with subject information; b) pipeline launcher; c) resource folder containing parametric images, report files and annotated information.
Keywords: image processing, preclinical imaging, XNAT, MR imaging, DICOM
203

Standardization and traceability of FDG PET imaging may improve statistical analysis (#470)

Thulaciga Yoganathan1, 2, Amel Raboudi2, 3, 4, Lucile Offredo1, 2, Thomas Viel1, 2, Bertrand Tavitian1, 2

1 Université Paris Descartes, Sorbonne Paris Cité, PARIS , France
2 INSERM U970, Paris-Cardiovascular Research Center at HEGP, PARIS , France
3 FEALINX, Courbevoie, France
4 Université de Technologie de Compiègne (UTC), UMR 7337 Roberval, Compiègne, France

Introduction

Standardization of animal handling, data acquisition and image reconstruction is essential for the ethical use of animals in research and impacts quantification, accuracy and reproducibility of small animal 18F-FDG Positron Emission Tomography (FDG PET) (1,2,3). In theory, using standardized PET acquisitions and image analysis, FDG uptake in healthy animals could be reused for multiple studies that include a similar control group. The present study aimed at testing whether data traceability and standardization can indeed help to minimize the number of mice in PET imaging.

Methods

A group of n=7 (TTT group) sunitinib-treated (50 mg/kg p.d. during 2 weeks) C57BL/6 male mice were compared to (i) an untreated control group (CON, n=9) from the same study and (ii) an untreated reference group (REF, n=52) of mice from our in-house database. Non-fasted mice were anesthetized with isoflurane, injected intravenously with FDG and imaged on the NanoScan PET-CT (Mediso, Hungary). Body weight, glycaemia, isoflurane and respiration were monitored. PET scans were reconstructed using customized 3D-OSEM and quantified using PMOD (Switzerland) by the same examiner. Mean comparison tests were performed between all groups using the Student test for groups of equal variances according to the Fisher test or the Wilcoxon-Mann-Whitney non parametric test at a 5% significance level.

Results/Discussion

Two weeks of sunitinib treatment increased FDG uptake in heart and muscle by 8% and 19% respectively in the TTT vs. CON group. However, the difference did not differ significantly between the 2 groups in the heart (SUVmean 6.33±0.96 vs. 5.85±1.51; p=0.4609) nor in the muscle (0.41±0.07 vs. 0.34±0.09; p=0.1059). In contrast, comparing the TTT and the REF groups demonstrated a significant difference in FDG uptake in the heart (p=0.0469) as well as in the muscle (p=0.0032). Access to a standardized PET database allows to evidence statistically significant differences in cardiac and muscle uptake induced by sunitinib treatment that we had failed to evidence in the original study because of the small number of mice used. Based on these encouraging results, we have implemented our PET-FDG database and the research workflow of standardized PET acquisitions, reconstructions and analyses, in a data management technology using the BMI-LM data model (4) in order to continuously feed the database.

Conclusions

The 52 PET exams of our current database are continuously fed and will now be extended to other acquisition conditions (e.g. fasting, anesthesia, sex, strain…). In the long term, such databases are expected to improve the observance of the 3R principle and allow power calculations in order to estimate the sample size needed to obtain a statistically significant result.

References

  1. Mannheim, J. G., Kara, F., Doorduin, J., Fuchs, K., Reischl, G., Liang, S., ... & Huisman, M. C. (2017). Standardization of Small Animal Imaging—Current Status and Future Prospects. Molecular Imaging and Biology, 1-16.
  2. Vanhove, C., Bankstahl, J. P., Krämer, S. D., Visser, E., Belcari, N., & Vandenberghe, S. (2015). Accurate molecular imaging of small animals taking into account animal models, handling, anaesthesia, quality control and imaging system performance. EJNMMI physics, 2(1), 31.
  3. Nanni, C., Rubello, D., & Fanti, S. (2007). Role of small animal PET for molecular imaging in pre-clinical studies.
  4. Allanic, M., Hervé, P. Y., Pham, C. C., Lekkal, M., Durupt, A., Brial, T., ... & Joliot, M. (2017). BIoMIST: A platform for Biomedical Data lifecycle Management of neuroimaging cohorts. Frontiers in ICT, 3, 35.

Acknowledgement

In vivo imaging was performed at the Life Imaging Facility of Paris Descartes University (Plateforme Imageries du Vivant - PIV), supported by France Life Imaging (grant ANR-11-INBS-0006), Infrastructures Biologie-Santé (IBISA), CARPEM Siric grant and Plan Cancer (ASC16026HSA-C16026HS).

Improved statistical significance using a standardized reference database

(A) Glycaemia, Body weight and injected dose in REF, CON and TTT groups: no difference by one-way ANOVA test (glycaemia and injected dose) and nonparametric Kruskal-Wallis test (body weight).

(B) FDG uptake was significantly increased in TTT heart and muscle using the REF but not the CON group for comparison. Heart *: p=0.047; £: p=0.223; $: p=0.461 Muscle **: p=0.007; ££: p=0.311; $$: p=0.207

 

 

Keywords: FDG PET imaging, Standardization, Traceability, Statistical analysis
204

Integrating PET-CT imaging research workflow with end-to-end traceability for small animal research (#337)

Amel Raboudi1, 2, 3, Thulaciga Yoganathan1, Thomas Viel1, Marianne Allanic2, Daniel Balvay1, Bertrand Tavitian1

1 INSERM, UMR970, Paris-Cardiovascular Research Center at HEGP, Paris, France
2 FEALINX, Courbevoie, France
3 Université de Technologie de Compiègne (UTC), UMR 7337 Roberval, Compiègne, France

Introduction

Poor annotation of heterogeneous data is a barrier to its reuse. The data model of Biomedical Imaging-Lifecycle Management (BMI-LM) platform enables data reuse. It provides provenance and domain concepts, as shown previously in neuroimaging studies [1]. Here, we assume that BMI-LM data model could be applied to PET-CT imaging research lifecycle and we propose a BMI-LM model-driven integration of PET-CT research workflow. Our data of interest are PET images in DICOM standard format, acquired since 2015 on a nanoScan PET-CT (Mediso, Hungary) during different small animal imaging studies.

Methods

BMI-LM is applied to COS-TEP: a retrospective research workflow using preclinical PET images. COS-TEP uses data from several research studies acquired in many configurations by different operators. The proposed PET-CT research workflow integration method covers (i) study specification (ii) raw data acquisition (iii) derived data analysis and (iv) publication (fig.1). First, BMI-LM provenance concepts are used to predefine and describe the final data. Then, acquired and reconstructed DICOM data integration is performed on-demand and accordingly PMOD analysis results are integrated [2]. Finally, BMI-LM database export to Excel is used for statistical analysis. The resulting integrated research workflow is based on a systematic end-to-end traceability using BMI-LM concepts.

Results/Discussion

The BMI-LM data model originally proposed for neuroimaging studies can be applied to PET-CT imaging research workflow provided that the following adaptations are made: (i) three new concepts are required, besides the BMI-LM existent ones[1], in order to manage PET-CT imaging data heterogeneity: Agent, Intervention and Sample. Agent includes description radiotracer and products used during the acquisition. Intervention designates the pharmacological treatment, the animal model preparation,..etc. And Sample describes the use of animal derived materials for in vivo and ex-vivo studies. (ii) For fluent adoption of the proposed method by imaging researchers, it is necessary to adapt the DICOM data integration tool to the local vocabulary of a PET preclinical imaging laboratory. Therefore, alignments between BMI-LM concepts, local vocabulary terms and DICOM standard were established and a Matlab GUI tool was provided. This assures more flexibility while maintaining DICOM standard compliance.

Conclusions

The ongoing integration of preclinical PET-CT imaging research workflow allows its continuous use for newly acquired data. Here, we have pointed some specific requirements of this integration. In future work, we intend to simplify it for PET-CT researchers, especially the study specification step bottleneck, using a customized Data Management Plan (DMP). This will allow scientists to reuse their own research workflow once it has been established.

References

1. Allanic, M. et al.: BIOMIST: A Platform for Biomedical Data Lifecycle Management of Neuroimaging Cohorts. Frontiers in ICT. 3, (2017).

2. Raboudi, A. et al.: Traçabilité de l’intégration de données biomédicales hétérogènes dans le système SWOMed de gestion du cycle de vie des études biomédicales. In: actes du symposium SIIM 2017. , Toulouse (2017).

Acknowledgement

This work is supported by the DRIVE-SPC collaboration project between Fealinx Company and LRI-PARCC-Inserm with a grant from IDEX Université Paris Sorbonne Cité. Amel Raboudi has an ANRT PhD scholarship. Authors want to thank PIV platform, Dr. Pierre-Yves Hervé, Dr. Philippe Boutinaud, Ing.Thierry Brial, Ing. Arthur Grioche, Ing. Olivier Menuel and Ing. Jérôme Cornet for their time and help.

The BMI-LM platform showing in flows and out flows with corresponding interfaces and concepts
This figure presents the proposed model-driven workflow integration for COS-TEP research workflow. Data integration tool was previously presented in [2] (in french). Presenting each BMI-LM concept relevant to COS-TEP adds complexity to the figure, that's why, groups of concepts are present instead. Please refer to [1] for detailed information about BMI-LM data model.
Keywords: Data provenance, research workflow, small animal research, PET-CT imaging, model-based workflow
205

Dispersion Correction for Sampled Blood Input function in Rats (#589)

Marie-Claude Asselin1, Daniela Bochicchio1, Rainer Hinz1, Hervé Boutin2

1 University of Manchester, Division of Informatics, Imaging & Data Sciences, Manchester, United Kingdom
2 University of Manchester, Division of Neuroscience & Experimental Psychology, Manchester, United Kingdom

Introduction

Tracer kinetic modeling of dynamic PET images requires an input function. Preclinically, blood-derived input functions suffer from severe dispersion in the narrow bore tubing connecting the animal to the blood sampler. Dispersion depends on the sampling rate, tubing material, inner diameter and length to and within the detector, and the dispersion correction must be derived for each experimental set-up and radiotracer. The aim of the study was to assess the accuracy of the dispersion correction proposed by Munk et al [1] as implemented in PMOD for use with the Twilite detector in rodents [2].

Methods

Solutions of approx. 1.2 MBq/mL were prepared by adding [F-18]FDG to water or blood. The dispersion function of PE tubing of 0.58-mm inner diameter and varying lengths (30, 45 or 60 cm) was measured by transferring the tubing from a beaker containing the non-radioactive solution to another beaker with the radioactive solution and back, creating a step function. Another 15 or 30 cm of tubing was winded in the F-18 or C-11 guide inserted into the Twilite detector. At the recommended sampling rate of 0.350 mL/min, the transit time in the detector is 6.8 s in the F-18 or 13.6 s in the C-11 guide. The sampling rate was also halved and doubled. The transmission-dispersion model [1], which does not account for the transit time in the detector, was used to fit the rising part of the step function

Results/Discussion

As shown in Fig. 1A, doubling and halving the sampling rate increased and decreased the slope of the rising step function. At the recommended sampling rate, the slope became shallower by lengthening the tubing to the detector (Fig. 1B) or inside the detector by replacing the F-18 guide with the C-11 guide (Fig. 1C). Substituting water with blood also led to shallower slope, to a similar extent an extra 15-cm of tubing (Fig. 1D). The Munk model did not fit the Twilite data because it was unable to follow the shape of the rising step function. The fitted parameters (alpha and kappa) were highly dependent on the initial values, particularly the transit time to the detector. For comparison, Munk et al [1] validated their model in pigs using the Allogg blood sampler operated at 7 mL/min and connected to PVC tubing of 1.65-mm inner diameter and 38-cm in length plus 5-cm inside the detector, corresponding to a transit time of only 0.9 sec.

Conclusions

By winding the tubing into the guide inserted in the Twilite detector in order to increase detection sensitivity, the Twilite detector acts as an integrator of the radioactive solution slowly transiting through it. The Munk model, which assumes instantaneous measurement of radioactivity, fails to fit the data and its parameters cannot be used to accurately correct the sampled blood input function for dispersion.

References

[1] Munk O.L. et al (2008) Med Phys 35(8): 3471-81.

[2] Alf M.F. et al (2013) J Nucl Med 54(1): 132-38.

Acknowledgement

To the EPRSC for contributing to the purchase of the Twilite blood sampler and to GSK for funding the PhD studentship of D. Bochicchio.

Figure 1
Effects of changing A) sampling rate, B) tubing length to the detector, C) guide and D) fluid (from water to blood) on the step function characterising the dispersion in the PE tubing connected to the Twilite detector.
Keywords: Blood input function, dispersion correction, tracer kinetic modeling
206

Low rank approximation as an alternative to accumulation for noise reduction in hyperpolarized 13C spectroscopy:  preliminary results for synthetic, phantom and in vivo spectra. (#560)

Roberto Francischello1, 2, Alessandra Flori3, 4, Luca Menichetti2

1 University of Pisa, Department of chemistry and industrial chemistry, Pisa, Italy
2 CNR, Clinical Physiology Institute, Pisa, Italy
3 Fondazione CNR/Toscana Gabriele Monasterio, Bioengineering and clinical engineering, Pisa, Italy
4 Scuola Superiore Sant'Anna, Institute of Life Sciences, Pisa, Italy

Introduction

Dissolution dynamic nuclear polarization allows in-vivo studies of metabolic flux using hyperpolarized 13C tracer by enhancing signal intensity.

To reduce the noise intensity and further enhance the signal-to-noise ratio Brender, et all. used low rank approximation on the real part of phase corrected or magnitude spectra for hyperpolarized 1-13C-pyruvate (13C-Pyr) [1,2].

Since the effect of the phase manipulation could alter the quantification of metabolites in these studies, we adapted the method of low rank approximation to be applied to complex value signal.

Methods

Let A be the NxM signal matrix which rows are made by the N acquired signal of n metabolites with n<N. Since the presence of additive noise makes A full rank it is possible to reduce the noise by searching for a low rank approximation of A.

The truncation of the last N-k singular value, obtained using singular value decomposition, gives the best approximation in Frobenius norm with rank k for A. The choice of k is related to the a-priori knowledge on the number of metabolites and to the variation of singular value.

We tested this method on simulated data and applied to panthom and in-vivo dataset. The phantom was made after dissolution of hyperpolarized 13C-Pyr in a standard volume. The animal model of rodent was injected with hyperpolarized 13C butyrate as previously reported [3].

Results/Discussion

To estimate the noise intensity, we take the standard deviation of the last points of the time-domain signal.

As shown in fig.1 the presented method is able to reduce the noise intensity in all the dataset, allowing the retrieval of signal even in extremely noisy data without the needing of phase pre-processing.

For the in-vivo sample the presence of a peak at ≈2 ppm was known from previews analysis, Gaussian apodization and sum of magnitude spectra [3]. Using low rank reduction was possible to observe a peak in the same spectral region, while retaining temporal resolution and spectral resolution thanks to a lower level of Gaussian apoditazion fig.2.

Conclusions

Our preliminary result shows that low rank approximation by SVD on complex signal matrix could be considered a valid alternative to accumulation of spectra for enhancing signal to noise ratio preserving temporal resolution in metabolic flux spectroscopy without the need of previous phase manipulation.

References

[1] KISHIMOTO, Shun, et al. Distinguishing Closely Related Pancreatic Cancer Subtypes In Vivo by 13C Glucose MRI without Hyperpolarization. BioRxiv, 2019, 511543

[2] BRENDER, Jeffrey R., et al. PET by MRI: Glucose Imaging by 13C-MRS without Dynamic Nuclear Polarization by Noise Suppression through Tensor Decomposition Rank Reduction. bioRxiv, 2018, 265793.

[3] FLORI, Alessandra, et al. Biomolecular imaging of 13C-butyrate with dissolution-DNP: Polarization enhancement and formulation for in vivo studies. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2018, 199: 153-160.

Acknowledgement

The author would like to thank the Fondazione CNR/Regione Toscana ‘G. Monasterio’ (Pisa, Italy) for funding this study.

Fig.1: Effect of denoise alghorithem on different dataset

Panel A): comparison between the spectra before and after the noise reduction for the 3 dataset. Panel B): comparison between the noise intensity before and after the noise reduction.

Fig.2: Comparison between magnitude spectra of butyrate
Comparison between the results of different post-processing technique on 13C butyrate. The ‘Sum’ spectrum was obtained by summing 20 Gaussian apodized (10Hz) magnitude spectra. The ‘SVD’ is the magnitude of the SVD low rank approximation of a single Gaussian apodized(1.5Hz) spectra.
Keywords: Hyperpolarization, Nuclear magnetic spectroscopy, Noise reduction, Low rank approximation
207

SPILL-IN EFFECTS IN POSITRON EMISSION TOMOGRAPHY (PET) IMAGING OF ABDOMINAL AORTA ANEURYSM (AAA) (#557)

Mercy I. Akerele1, Daniel Deidda1, 2, Jacobo Cal-Gonzalez3, Rachael O. Forsythe4, Marc R. Dweck4, Nicolas A. Karakatsanis5, 6, Robert G. Aykroyd2, Steven Sourbron1, Charalampos Tsoumpas1, 6

1 University of Leeds, Biomedical Imaging Science Department, Leeds, United Kingdom
2 University of Leeds, Department of Statistics, Leeds, United Kingdom
3 Medical University of Vienna, Centre for Medical Physics and Biomedical Engineering, Vienna, Wien, Austria
4 University of Edinburgh, Centre for Cardiovascular Science, Edinburgh, United Kingdom
5 Weil Cornell Medical University, Department of Radiology, New York, New York, United States of America
6 Icahn School of Medicine, Translational and Molecular Imaging Institute, New York, New York, United States of America

Introduction

Positron emission tomography (PET) is an imaging tool with primary applications in oncology, neurology and cardiology. However, accurate clinical assessment is often affected by the partial volume effect (PVE), leading to overestimation (spill-in) or underestimation (spill-out) of activity in various small regions. Many techniques are used to correct for PVE, but past studies have shown that most of these techniques cannot effectively correct for the spill-in effects when the target region is within 1-5 cm to a highly radioactive region such as the urinary bladder, myocardium or bone.

Methods

This study evaluated the spill-in effects using 3 patients [18F]-Sodium Fluoride PET data of abdominal aortic aneurysms (AAA) adjacent to active bone, acquired with the Siemens Biograph mCTTM scanner. We also compared the performance of various recently developed spill-in correction techniques, namely: (1) background correction (BC), (2) local projection (LP), and (3) hybrid kernel method (HKEM) methods. Iterative reconstructions, including point spread function (PSF) modelling, were performed with the Software for Tomographic Image Reconstruction (STIR) library. Region of Interest (ROI) analysis was performed by comparing the SUVmax and target-to-background ratio (TBR) of two delineating ROIs: one covering the entire aneurysm and the other excluding the parts very close to the bone.

Results/Discussion

The results demonstrated large differences in corrected SUVmax and TBRmax scores between the ROIs drawn over the entire aneurysm (AAA) and ROIs excluding some regions close to the bone (AAA-exc) for both cSUVmax and TBRmax especially for patient 3. This discrepancy in SUV values between the two ROIs is likely due to the spill-in effect emanating from the adjacent active bone. LP and HKM performed well in reducing this spill-in activity for patients 1 and 2, but not for patient 3. PSF+BC had the least difference for all 3 patients under review, suggesting the best spill-in correction performance. These results show that the uptake measurement in the abdominal aneurysm is highly dependent on how the ROI is drawn, and most importantly on clinician expertise. In addition, excluding some regions close to the bone in a bid to avoid the spill-in effect from the bone, could result in a certain degree of potentially important physiological information being lost from the excluded regions.

Conclusions

Overall, the BC technique yielded the best performance in spill-in correction for the patient data. This implies that BC can be used to effectively correct the spill-in effect from the bone into the aneurysm. Additionally, BC method is also robust to ROI-selection variability and could thereby enhance accurate quantification in regions of interest.

References

[1] Y. Liu. “Invalidity of SUV measurements of lesions in close proximity to hot sources due to shine-through effect on FDG PET-CT interpretation,” Radiol. Res. Pract. 868218, 2012.

[2] M. I. Akerele, P. Wadhwa, J. Silva-Rodriguez, W. Hallett and C. Tsoumpas, “Validation of the physiological background correction method for the suppression of the spill-in effect near highly radioactive regions in positron emission tomography,” EJNMMI Phys., 5:34, 2018.

[3] R. O. Forsythe, M. R. Dweck, O. M. B. McBride, A. T. Vesey, S. I. Semple, A. S. V. Shah, et al., “18F-Sodium Fluoride Uptake in Abdominal Aortic Aneurysms: The SoFIA3 Study,” J. Am. Coll. Cardiol., vol. 71, no. 5, pp. 513-523, 2018.

[4] J. Cal-Gonzalez, X. Li, D. Heber, I. Rausch, S. C. Moore, K. Schafers, et al., “Partial volume correction for improved PET quantification in 18F-NaF imaging of atherosclerotic plaques,” J. Nucl. Cardiol., vol. 25, no. 5, pp. 1742-1756, 2017.

[5] D. Deidda, N. A. Karakatsanis, P. M. Robson, N. Efthimiou, Z. A. Fayad, R. G. Aykroyd, and C. Tsoumpas, “Effect of PET-MR inconsistency in the kernel image reconstruction method,” IEEE Trans. Radiat. Plasma Med. Sci, 2018 (in press).

Acknowledgement

This work was undertaken on MARC1, part of the High-Performance Computing and Leeds Institute for Data Analytics (LIDA) facilities at the University of Leeds, UK.

Corrected SUVmax (cSUVmax) and maximum target-to-backgriound ratio (TBRmax)

Fig. 1. The cSUVmax and the TBRmax analysis from patients 1 - 3 (left-to-right columns), as estimated for all evaluated reconstruction algorithms at 3 iterations and with 3 mm Gaussian post-filter.

Reconstructed Images from all correction techniques

Fig. 2. Sagittal views of the PET reconstructed images at 3 full iterations with 3 mm Gaussian post-filter, shown for all 3 patients (top-to-bottom rows). The ROIs used to extract the SUVs for AAA (outer sphere) and AAA-exc (inner sphere) are shown on the CTAC image.

Keywords: PET, PVE, spill-in effect, quantification
208

Biological diversity or numeric artifact? MSI data clustering robustness. (#587)

Grzegorz Mrukwa1, Katarzyna Fratczak1, Marta Gawin2, Mykola Chekan2, Monika Pietrowska2, Piotr Widlak2, Joanna Polanska1

1 Silesian University of Technology, Data Mining Group, Gliwice, Poland
2 The Centre of Oncology -Maria Skłodowska-Curie Institute, branch in Gliwice, Gliwice, Poland

Introduction

Mass Spectrometry Imaging is a recent technique providing vast amounts of data about the distribution of hundreds of compounds within a tissue, revealing implicit structure crucial for investigating tumor or neurodegenerative diseases. Despite dedicated preprocessing, data is heavily duplicated (thus overly multidimensional) and noisy. Moreover, in molecularly diverse regions, different descriptors play the key role in heterogeneity analysis. A number of clustering algorithms were presented, as well as dimensionality reduction techniques. But which combination provides justified results?

Methods

Data from two Head and Neck Cancer cases (8005 and 11 869 spectra, 109 568 mass channels each) was subject to the standard pipeline of Savitzky-Golay smoothing, Peak Alignment via Fast Fourier Transform, baseline removal, and TIC normalization. Instead of overly simplistic procedures for peak selection, peak modeling via a Gaussian Mixture Model (GMM) was applied and 3 714 new descriptors were obtained.

Three clustering algorithms were included in considerations: standard K-Means algorithm, graph-cuts based clustering, and Divisive intelligent K-Means. Each of them ran independently on GMM-modelled data, PCA-reduced dataset (most relevant components w.r.t. knee-method or EXIMS score) and dataset reduced by nonlinear manifold embedding. The number of clusters was indicated by GAP index.

Results/Discussion

From each scenario, clusters were selected to compose regions of tumor and epithelium (figure). Clusters from the graph-cuts deformed region of tumor and epithelium marked by pathologist and exhibit fragmentation over homogeneous regions of the epithelium.

K-means provides the most spatially consistent clusters (except EXIMS), but only with manifold embedding is capable to reveal major diversity within the healthy epithelium.

DiviK for GMM components reveals both tumor and epithelium molecular diversity. Results are stable for PCA transformation, despite small clusters appear (knee) or vanish (EXIMS). With manifold embedding provides a similar result to k-means (manifold), and, similarly to the DiviK (GMM), diversity within the cancer region is captured.

The effect size of features was computed and features with at least large effect size between clusters were measured. Top 3 counts correspond to DiviK algorithm with PCA (knee), GMM and manifold. Most of the counted features are the same.

Conclusions

Inspection of clusters indicates three candidates for biological diversity analyses: K-means and DiviK after manifold embedding, as well as DiviK in GMM space. PCA (dependent on the component selection method) tends to lose or exaggerate nuances in the data. Spectral clustering introduces fragmentation that is hardly corresponding to tissue diversity. Finally, DiviK finds regions which could be differentiated with the greatest number of features.

References

  1. Dexter, Alex, et al. "Two-phase and graph-based clustering methods for accurate and efficient segmentation of large mass spectrometry images." Analytical chemistry 89.21 (2017): 11293-11300.
  2. Pietrowska, Monika, et al. "Molecular profiles of thyroid cancer subtypes: Classification based on features of tissue revealed by mass spectrometry imaging." Biochimica et Biophysica Acta (BBA)-Proteins and Proteomics 1865.7 (2017): 837-845.
  3. Tibshirani, Robert, Guenther Walther, and Trevor Hastie. "Estimating the number of clusters in a data set via the gap statistic." Journal of the Royal Statistical Society: Series B (Statistical Methodology) 63.2 (2001): 411-423.
  4. Wijetunge, Chalini D., et al. "EXIMS: an improved data analysis pipeline based on a new peak picking method for EXploring Imaging Mass Spectrometry data." Bioinformatics 31.19 (2015): 3198-3206.
  5. Polanski, Andrzej, et al. "Signal partitioning algorithm for highly efficient Gaussian mixture modeling in mass spectrometry." PloS one 10.7 (2015): e0134256.

Acknowledgement

This project was financially supported by the European Union through the European Social Fund (grant POWR.03.02.00-00-I029) (GM) and NCN grant BITIMS no. UMO-2015/19/B/ST6/01736 (GM, JP, KF).

Optical scan of analyzed tissue
Tissue regions are marked over the preparation area. Red color corresponds to tumor tissue and inflammatory response. Cyan color corresponds to the healthy epithelium.
Visualization of clustering results
Clustering results are limited to the area corresponding to the tumor and healthy epithelium. Different colors indicate separate clusters. Colors were selected in a way that matches original H&E coloring the most.
Keywords: mass spectrometry imaging, tissue heterogeneity, clustering, dimensionality reduction
209

PSMA-PET/CT texture analysis to characterise unaffected bone and bone metastases of patients with prostate cancer (#408)

Robert Seifert1, 2, Aaron Scherzinger3, Bastian Zinnhardt1, 2, Florian Büther1, Matthias Weckesser1, Kambiz Rahbar1, Michael Schäfers1, 2

1 University of Münster, Department of Nuclear Medicine, Münster, North Rhine-Westphalia, Germany
2 University of Münster, European Institute for Molecular Imaging, Münster, North Rhine-Westphalia, Germany
3 University of Münster, Department of Computer Science, Münster, North Rhine-Westphalia, Germany

Introduction

Bone metastases of prostate cancer can be detected by various modalities like bone scintigraphy, CT or PSMA PET acquisitions. Yet the characterization of Disseminated Tumour Cells (DTCs) in the bone marrow by imaging means is an unmet clinical challenge. Texture analysis (TA) is an established tool to reveal image features of clinical significance and has recently been applied to PSMA PET acquisitions [1]. TA is often utilised for the analysis of tumours or metastases. Here, TA is used to characterise bone metastases and apparently unaffected bone parts, i.e. with no visible bone metastases.

Methods

98 [68Ga]-PSMA-11 PET/CT acquisitions were automatically analysed (25 without, 73 with bone metastases). CT acquisitions were used to segment a bone mask for each PET/CT, which was then applied to the PET acquisitions to exclude non-bone structures. Further, a global SUV threshold was defined to segment bone metastases in each PET acquisition. To cope for potential spill-out artefacts, each binary bone metastases mask was expanded by a Gaussian smoothing of 8.6 mm FWHM (thus minimising effects of positron ranges and system resolution) and thresholded to include 99.9 % of the Gauss distribution. Fifteen texture analysis features were calculated for bone metastases and unaffected bone of each PET acquisition, including Grey Level Cooccurrence Matrix (GLCM) features [2].

Results/Discussion

The global threshold to segment visible bone metastases was set to 1.6 SUV. In an exploratory approach, Receiver Operating Characteristics (ROC) were calculated for each textural feature to assess the performance of unaffected bone TA to classify acquisitions as either bone metastases free or osseous metastasized. Highest Area Under Curve (AUC) value was obtained by the GLCM feature ‘angular second moment’ (GLCMASM; AUC = 0.711), whereas the best statistical image feature was ‘image entropy’ (AUC = 0.334; for comparison AUC of SUVmean: 0.352) [2]. Analysing unaffected bone, there was a significant difference comparing GLCMASM of patients with and without bone metastases (0.013 vs. 0.002, respectively; p < 0.005). However, TA of bone metastases seemed of little clinical relevance: PSA levels correlated with SUVmean but not with GLCMASM of bone metastases (Spearman: p < 0.001 vs. p > 0.05).

Conclusions

Automated TA of visually unaffected bone is feasible and reveals image features that were significantly different between PET acquisitions with and without bone metastases. TA of bone metastases seems inferior to SUVmean. Future studies have to show if image features of unaffected bone have a predictive value for the outcome and if they correlate with the presence of DTCs.

References

  1. Khurshid Z, Ahmadzadehfar H, Gaertner FC, et al (2018) Role of textural heterogeneity parameters in patient selection for 177Lu-PSMA therapy via response prediction. Oncotarget 9:33312–33321. https://doi.org/10.18632/oncotarget.26051
  2. Haralick RM, Shanmugam K, Dinstein I (1973) Textural Features for Image Classification. IEEE Trans Syst Man Cybern SMC-3:610–621

Acknowledgement

none.

Keywords: PSMA PET/CT, texture-analysis, bone metastases, prostate cancer.
210

Simplified estimation of binding potential for TSPO tracer 18F-DPA714 without a reference region (#236)

Claire Leroy2, Sonia Lavisse2, 4, Martin Schain3, Michel Bottlaender2, 5, Irène Buvat2, Catriona Wimberley1, 2

1 University of Edinburgh, Edinburgh Imaging Facility QMRI, Edinburgh, United Kingdom
2 CEA, Inserm, Université Paris-Sud, Université Paris Saclay, IMIV, Orsay, France
3 Copenhagen University Hospital, Neurobiology Research Unit, Copenhagen, Denmark
4 CEA and CNRS-UMR9199, Université Paris-Sud, MIRCen, Fontenay-aux-Roses, France
5 CEA, Neurospin, Gif-sur-Yvette, France

Introduction

Quantification of TSPO tracers such as 18F-DPA714 faces several challenges1 including lack of a reference region. Arterial blood sampling is gold standard, but not always possible so alternative quantification methods are valuable. There are BPND estimation methods for 18F-DPA714 such as supervised cluster analysis to extract a reference curve (SVCA)2,3 however, these methods are susceptible to bias from whole brain TSPO. The aim of this study is to assess a previously developed method of estimating BPND (SIME) 4 on 18F-DPA714 scans using an individual AIF and a template input function (tIF).

Methods

Ten healthy subjects (3 mixed (MAB), 7 high affinity binders (HAB)) underwent 18F-DPA-714 scans with metabolite corrected arterial input functions (AIF). For each subject, SIME analysis used the 2TCM with a fixed VND across 15 brain regions, fitting all regions simultaneously4. A grid of 100 VND values was tested (0.1 to 10) and the VND that produced the lowest cost function was selected. The regional BPNDs were calculated using the VTs and VND. This was done with two input functions and compared:

  1. Individual AIF
  2. A tIF was created for each subject by averaging the AIFs: each AIF was normalised (area under the curve=1) and the peaks temporally shifted so they were lined up with each other. For each subject, a tIF was generated from the 9 other subjects, to emulate when there is no AIF

Results/Discussion

The SIME method successfully identified VND values for each subject using its own AIF. Mean VND values were HABs: 2.7±1.30 ml/cm3 and MABs: 2.0±0.06 ml/cm3 which were used to estimate regional BPND as shown in Fig 1. The BPND estimates using the tIF were strongly correlated with those using individual AIFs in HAB and MAB groups: r2 = 0.98 and 0.95 respectively (Fig 2).

To assess the value of the method, it could be tested in a cohort of pathological subjects and compared against SUVr or BPNDs calculated using reference tissue models with the cerebellum or extracted SVCA curve. This would allow us to see if SIME is more sensitive to regional changes in TSPO than the other methods, which could be susceptible to bias from TSPO expression across the whole brain or unexpected regional increases in TSPO expression. On top of that, the SIME method could be tested with a model including a component for vascular TSPO.

Conclusions

We have applied the previously developed SIME method of estimating BPND to 18F-DPA714, which identified BPND values using individual AIFs and also the population based tIF. The BPND values between the two methods were highly correlated. The SIME method combined with a tIF will be useful for obtaining regional BPND estimates without individually sampled AIFs or a reference region.

References

  1. Turkheimer et al.; The methodology of TSPO imaging with positron emission tomography; Biochem. Soc. Trans., 2015
  2. Garcia Lorenzo et al.; Validation of an automatic reference region extraction for the quantification of [18F]DPA-714 in dynamic brain PET studies; JCBFM, 2017
  3. Turkheimer et al.; Reference and target region modelling of 11C-R-PK11195 brain studies; JNM, 2007
  4. Schain et al.; Non-invasive estimation of 11C-PBR28 binding potential; Neuroimage, 2018

Acknowledgement

Dr Catriona Wimberley's position as Research Fellow in PET-MRI physics at the University of Edinburgh is funded by Siemens Healthcare Limited, and her previous post doctoral position was supported by a CEA-Enhanced Eurotalents cofund with FP7 Marie Sklodowska-Curie COFUND Program (600382).

Regional BPND estimates

Average regional BPND estimates for HABs and MABs using the SIME 2TCM with individual AIFs

Correlation between AIFs and tIFs.

Correlations between regional BPND estimates calculated using the individual AIFs and the tIF for HABs and MABs

Keywords: TSPO, quantification, kinetic modelling, 18F-DPA714, simplified
211

Segmentation of amyloid-β plaques in three mouse models of Alzheimer’s disease using X-ray phase contrast-computed tomography (#362)

Coralie Gislard1, Carlie Boisvert1, Cécile Olivier2, Françoise Peyrin2, Marlène Wiart3, Hervé Boutin4, Hugo Rositi5, Fabien Chauveau1

1 Univ. Lyon, Lyon Neuroscience Research Center; CNRS UMR5292; INSERM U1028, Univ. Lyon 1, Lyon, France
2 Univ.Lyon, CREATIS; CNRS UMR5220; INSERM U1044; INSA-Lyon; Univ. Lyon 1, Lyon, France
3 Univ. Lyon, CarMeN laboratory; INSERM U1060; INRA U1397; Hospices Civils de Lyon, Lyon, France
4 Univ. Manchester, Faculty of Biology Medicine and Health, Wolfson Molecular Imaging Centre, Manchester, United Kingdom
5 Univ. Clermont Auvergne; CNRS; SIGMA Clermont; Institut Pascal, , Clermont-Ferrand, France

Introduction

X-ray Phase Contrast Tomography (XPCT) uses highly coherent synchrotron radiation to image soft tissues [1]. Ex vivo brain XPCT enables a virtual histology of cerebral structures [2], myelinated tracts [3], but also amyloid plaques (Aβ) [4]. Previous reports on Aβ detection have been mostly restricted to qualitative observations [5], owing to the difficulty to process the large amount of data arising from whole-brain imaging at a µm scale. The present work aims to develop a segmentation pipeline to extract relevant measurements on Aβ plaques across several Alzheimer’s transgenic lines.

Methods

Three transgenic lines were used, for a total of eight brains: i) 3xTg (n=3, 1 y.o.), ii) APP-PS1 (n=3, 1 y.o.), iii) PDAPP-J20 (n=2, 2 y.o.) [6].

Fixed brains were dehydrated in ethanol and placed in test tubes for imaging at ESRF beamline ID-19 (Table 1). Paraffin embedding and standard immunofluorescence using 4G8 antibody and thioflavin S (ThS) were performed afterwards.

Semi-automated detection of Aβ plaques in the hippocampus used Fiji software and the following plugins: segmentation editor (to isolate hippocampus), trainable WEKA segmentation (to identify plaques) and 3D objects counter (to extract relevant parameters). Accuracy of segmentation was visually evaluated.

Results/Discussion

The semi-automated detection pipeline identified, in the hippocampus of a single hemisphere, 25±5 plaques in 3xTg, 1350±377 plaques in APP-PS1 and 2577±457 plaques in PDAPP-J20, which presence was confirmed by immunofluorescence (Fig. 1). No false-positive structures, such as hippocampal neurons, were segmented. The mean volume size of individual plaques in 3xTg was about twice the ones of APP-PS1 and PDAPP-J20 (25470±4360 vs 14150±4360 vs 13650±1720 µm3 respectively), and the fraction volume of hippocampus occupied by Aβ was 0.01% in 3xTg, 0.22% for APP-PS1, and 0.48% for PDAPP-J20. In all cases, the majority of plaques were ellipsoidal, with 15 to 40% being detected as circular.

Conclusions

A pipeline using freely available tools was successfully tested on three transgenic lines, and accurately detected visible Aβ plaques. The present work leveraged 3D-analysis to extract quantitative parameters hardly available with standard histology. The validation of such quantification pipelines is a necessary step to realize the full potential of XPCT in neuroscience research.

References

[1] Albers et al. Mol Imaging Biol. 2018.

[2] Marinescu et al. Mol Imaging Biol. 2013; Barbone et al. IJROBP 2018.

[3] Rositi et al. In: 12h European Molecular Imaging Meeting (EMIM), Cologne, Germany. 2017.

[4] Chauveau et al. In: 13th European Molecular Imaging Meeting (EMIM), San Sebastian, Spain. 2018.

[5] Noda-Saita  et al. Neuroscience 2006 ; Connor et al. NeuroImage 2009 ; Pinzer et al. NeuroImage 2012 ; Astolfo et al. J Synchrotron Radiat. 2016 ; Massimi et al. NeuroImage 2019.

[6] Oddo et al. Neuron. 2003 ; Jankowsky et al. Hum Mol Genet. 2001 ; Mucke et al. J Neurosci. 2000.

Acknowledgement

MITACS grant (Carlie Boisvert); LABEX  PRIMES  (ANR-11-LABX-0063).

Table 1.
Acquisition parameters used for whole-brain imaging of mouse brains at beamline ID-19 of the European Synchrotron Radiation Facility (ESRF)
Figure 1.
Corresponding (A) XPCT image, (B) amyloid staining with 4G8 antibody, and (C) Thioflavin S staining, for the three transgenic mouse strains (I- 3xTg; II- PDAPP-J20; III- APP-PS1).
Keywords: X-ray Phase Contrast Tomography, amyloid, Alzheimer's disease, virtual histology
212

Effect of image-derived input function extraction method on quantification of 18F-FDG uptake in mice with different dietary conditions (#192)

Celia De La Calle3, Natalia Magro1, Marta Ibañez1, Marta Oteo1, Guillermo Garaulet2, Francisca Mulero2, Alejo Efeyan3, Miguel Angel Morcillo1

1 CIEMAT, Biomedical Applications of Radioisotopes and Pharmacokinetics Unit, Madrid, Spain
2 CNIO, Molecular Imaging Core Unit, Madrid, Spain
3 CNIO, Metabolism and Cell Signalling Lab, Madrid, Spain

Introduction

Measurement of an accurate plasma input function is required for quantitative 18F-FDG PET studies in mice. We aimed to compare the effect of method of generating image-derived input function (IDIF) on the quantification of 18F-FDG uptake in mice PET studies under different dietary states.

Methods

Mice (C57BL6x129S2/SvPasCrl) were studied under different conditions: nonfasted, fasted and insulin administration. 60-min dynamic PET was performed under isoflurane anesthesia after 18F-FDG injection. Regions of interest were manually drawn over: myocardium, liver, cerebral cortex, brown adipose tissue, muscle and white fat. IDIFs were obtained from the left ventricle cavity (LV) using the COMKAT software1 or vena cava (VC) fitted to the 3-compartment FDG model using nonlinear regression2. The 18F-FDG uptake constant (Ki) was estimated by Patlak graphical analysis3 using input function and time-activity data from 15-60 min. When the resulting plot became horizontal, a Logan plot was applied to calculate the distribution volume4. Tracer uptake was also quantified as SUVs.

Results/Discussion

Arterial time-activity curves (TAC) normalized by injected dose (SUV) are illustrated in Figure 1. No significant differences in the area under the arterial curve (AUC; expressed as SUV x min) for LV and VC were observed between nonfasted and fasted or insulin mice (333±189 vs. 301±126 and 413±26 for LV and 108±13 vs. 114±33 and 79±30 for VC). The AUCs were significantly higher in TAC obtained from LV cavity than from VC due the spillover from the myocardial wall and the partial-volume non-correction of the VC.  Result of the graphical analysis of PET data is shown in Table 1. No significant difference in the Ki estimates from myocardium and cerebral cortex was obtained with LV and VC IDIF, although the values were higher in those estimated from LV IDIF. The myocardium Ki values were significant higher in insulin than in nonfasted mice. There were no significant differences in distribution volume values obtained with both IDIFs in the rest of organs, irrespective of dietary state.

Conclusions

The arterial left ventricle and vena cava time-activity curves can be used both as an easy accessible input function for kinetic modeling of 18F-FDG uptake. Although a slight underestimation of Patlak/Logan slope from the vena cava as IDIF compared to the obtained from left ventricle was found, the choice of IDIF had not significant effect on Patlak/Logan kinetics and calculated 18F-FDG uptake under different dietary conditions.

References

1Muzic RF, Jr, Cornelius S. COMKAT: compartment model kinetic analysis tool. J Nucl Med. 2001;42:636–45. (http//comkat.case.edu)

2Feng D, Wong KP, Wu CM, Siu WC. A technique for extracting physiological parameters and the required input function simultaneously from PET image measurements: theory and simulation study. IEEE Trans Inf Technol Biomed. 1997;1:243–54.

3 Patlak CS, Blasberg RG. Graphical evaluation of blood-to-brain transfer constants from multiple-time uptake data. Generalizations. J Cereb Blood Flow Metab. 1985;5:584-590.

4 Logan J. Graphical analysis of PET data applied to reversible and irreversible tracers. Nucl Med Biol. 2000;27:661-670.

Figure 1. Time-activity curves
Time-activity curves averaged for nonfasted (n=4, black), fasted (n=4, red) and insulin-treated (n=3, blue). Given are blood activities divided by injected dose per Kg of body weight (SUV) for left ventricle cavity and vena cava.
Table 1. Graphical analysis

Ki (mL tissue/mL plasma*min) and DV (mL tissue/mL plasma) values estimated by graphical analysis using IDIFs obtained from left ventricle and vena cava. Mice: nonfasted (n=4), overnight fasted (n=4) and insulin-treated (n=3). (a) Significant differences were found between nonfasted and insulin-treated mice. (b) Significant difference was found between LV and VC IDIF for nonfasted mice.

Keywords: PET, quantitative image analysis, IDIF
213

Radiographic quantitative values measured from 18F-NaF PET/CT bone scan for Alkaptonuria patients at Lumbar spine and femoral region. (#158)

Eman H. Alawadhi1, Sobhan Vinjamuri2, James Gallagher1, Ranganath Lakshminarayn3

1 The University of Liverpool, Musculoskeletal Biology I, Institute of Ageing & Chronic Disease, William Henry Duncan Building, Liverpool, United Kingdom
2 Royal Liverpool University Hospital, Department Of Nuclear Medicine, Liverpool, United Kingdom
3 Royal Liverpool & Broad green University Hospitals , Liverpool Clinical Laboratories, Liverpool, United Kingdom

Introduction

Alkaptonuria is a genetioarthritic disorder caused by a deficiency of HGA enzyme leading to increasing homogentisic acid forming cartilage pigmentation and arthritis called ochronosis. Radiographic quantitative assessment from PET/CT using bone radiotracer can be used to assess the progression of skeletal involvement in AKU.  

The study aim is to assess bone involvement in lumbar and hip regions for AKU by measuring Hounsfield units from CT and standardised uptake value from PET. We aimed also to determin the correlation between HU, SUV and bone density value and tseting age and gender affect.

Methods

A total of 39 AKU patients (males 24, female 15) who involved in the SONIA2 trial and underwent the whole body 18F-NaF PET/CT and DEXA scans were enrolled in our study. The mean age of the patients was 50 ± 10, and each gender was grouped into four groups stratified by decade of life. A HERMES software system was used to measure an average HU from CT image and SUVmax from PET image for lumbar vertebrae L1-L5 and head of the femur. For each measurement, the largest possible elliptical region covering the ROI was drawn in the axial slice, excluding the cortical margins to avoid volume averaging (fig1).The effect of age in each gender for HU and SUVmax was tested. The measured quantitative values were compared between each other age and correlated with bone mineral density from a DEXA scan.

Results/Discussion

In our study, although that males have higher T-score values compared with females, there was no a significant difference in HU and SUVmax between genders. The mean HU across L1-L5 was 145 for males and 152 for female.HU values showed a significant correlation with increased age for both gender in average lumbar vertebral bodies (male, p= 0.014, r = 0.48; female p= 0.0005, r =0.79) but only for female in head of the femur with increasing age (r = 0.83 and p < 0.0005). SUV values were not changed significantly between the groups of age except in average lumbar vertebrae for males (r= 0.47, p= 0.021). The range of lumbar vertebrae SUVmax was between 5.2 to 15.91 while femoral SUVmax was between 2.2 to 21. There was no correlation coefficient of SUVmax with T score (p > 0.05) in both lesions, but moderate correlations with HU in lumbar lesion only. There was a significant correlation coefficient of HU and T score in lumbar and femur region (p < 0.005).

 

Conclusions

The present study proposes a novel radiographic quantitative method by measuring morphological and functional quantitative data from a single scan which is readily available but rarely used. SUV can be used as a quantitative method to quantify bone PET/CT studies which convey bone metabolism and structural information. HU values from PET/CT could be useful in measuring bone density based on the strong correlation between HU and bone density valu.

References

[1] W. J. Introne and W. A. Gahl, “Alkaptonuria,” 2016.

[2] P. Omoumi, G. A. Mercier, F. Lecouvet, P. Simoni, and B. C. Vande Berg, “CT Arthrography, MR Arthrography, PET, and Scintigraphy in Osteoarthritis,” Radiol. Clin. North Am., vol. 47, no. 4, pp. 595–615, 2009.

[3] N. Kobayashi et al., “New application of 18F-fluoride PET for the detection of bone remodeling in early-stage osteoarthritis of the hip.,” Clin. Nucl. Med., vol. 38, no. 10, pp. e379-83, 2013.

[4] N. Kobayashi et al., “Use of 18F-fluoride positron emission tomography as a predictor of the hip osteoarthritis progression.,” Mod. Rheumatol., vol. 25, no. 6, pp. 925–930, 2015.

Acknowledgement

The author specially thanks to Professor James Gallagher, Professor Sobhan Vinjamuri and Professor Ranganath Lakshminarayn for supporting and supervise me in my PhD research

PET/CT scan

Upper images, illustrating the methods of determining the HU values and SUV using an elliptical ROI. a. Shows the vertebral bodies in a sagittal slice of PET/CT of the lumbar SPINE. In the panel, three axial locations are selected b. immediately superior to inferior endplate c. middle vertebra and d. inferior to the superior endplate.

Lower images; axial slice of head of the femur.
Keywords: semiquantitative assessment, Alkaptonuria, standarised uptake value, Hounsfield Units, 18F-NaF PET/CT
214

An Open-Source Pipeline for Correlation and Visualization of Multi-Modal Multi-Scale Imaging Data in Tumor Vascularization (#114)

Verena Stanzl1, Lydia M. Zopf2, Stefan H. Geyer3, Nicole Swiadek4, Jelena Zinnanti2, Paul Slezak4, Wolfgang J. Weninger3, Andreas Walter5, Katja Bühler1

1 VRVis Zentrum für Virtual Reality und Visualisierung Forschungs-GmbH, Vienna, Austria
2 Vienna Biocenter Core Facilities GmbH, Preclinical Imaging Facility, Vienna, Austria
3 Medical University of Vienna, Division of Anatomy, Center for Anatomy and Cell Biology & MIC, Vienna, Austria
4 Ludwig Boltzmann Institute for Experimental and Clinical Traumatology, Vienna, Austria
5 BioImaging Austria - Correlated Multimodal Imaging Node Austria (CMI), Vienna, Austria

Introduction

Activating angiogenesis is one of the main indicators of cancer supporting rapid tumor growth and metastasis [1]. Current treatment strategies aim at inhibiting tumor neo-angiogenesis, but tumor vessels are disorganized, leaky and dysfunctional [2]. To understand how tumors instruct the formation and maintenance of blood vessels, we apply ultra-high field MRI, µCT and HREM to a xenograft mouse model and correlate the tumor vessels. The proposed open source visualization pipeline supports the analysis of angiogenesis by depicting the different vessels and anatomical context in a single model.

Methods

Vessel segmentation is performed using semi-automatic vessel tracking using [3] and Amira. The visualization pipeline provides an open source approach to integrate the multi-modal multi-scale imaging data. µCT is used as reference space for data integration. Its resolution is high enough to show enough detail from HREM data while providing a data size manageable by most software tools. Thus, all data is rescaled to µCT resolution using 3D Slicer [4]. We use CustusX [5] and its 3D visualization to perform pairwise affine landmark-based registration as spatial orientation is crucial to mark corresponding landmarks at branching points. The generated transformation matrices are used to resample all data to the µCT reference in 3D Slicer. Multi-volume rendering is performed with tomviz [6].

Results/Discussion

The visualization in our example (see Figure 1) shows the tumor's blood supply provided by surrounding vessels and reveals the finer vessels subserving the tumor itself. Due to different resolutions, the main branch visible in MRI is also detected in the other modalities, whereas smaller vessels can only be seen in µCT and HREM.

Currently affine image correlation is performed to establish correspondence of matching branching points. However, non-rigid registration would be necessary to correlate the vessels more accurate compensating for deformations the vessels undergo during preparation processes for image acquisition. This is an open problem even for state-of-the-art registration software, as corresponding landmarks along incomplete vessel segments are unknown and image context cannot easily be correlated.

Conclusions

We presented a processing pipeline for visualizing the tumor vascularization in a xenograft mouse model gained by MRI, µCT and HREM to support the analysis of angiogenesis.

The pipeline is still fragmented and the development of a tool providing an integrated correlation and visualization workflow remains an open issue and future work.

References

[1] Folkman, J. (2002). Role of angiogenesis in tumor growth and metastasis. In Seminars in oncology (Vol. 29, No. 6, pp. 15-18). WB Saunders.

[2] Cao, Y. (2008). Molecular mechanisms and therapeutic development of angiogenesis inhibitors. Advances in cancer research, 100, 113-131.

[3] Novikov, A. A., Major, D., Wimmer, M., Sluiter, G., & Bühler, K. (2017). Automated Anatomy-Based Tracking of Systemic Arteries in Arbitrary Field-of-View CTA Scans. IEEE transactions on medical imaging, 36(6), 1359-1371.

[4] Fedorov, A., Beichel, R., Kalpathy-Cramer, J., Finet, J., Fillion-Robin, J. C., Pujol, S., ... & Buatti, J. (2012). 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magnetic resonance imaging, 30(9), 1323-1341.

[5] Askeland, C., Solberg, O. V., Bakeng, J. B. L., Reinertsen, I., Tangen, G. A., Hofstad, E. F., ... & Hernes, T. A. N. (2016). CustusX: an open-source research platform for image-guided therapy. International journal of computer assisted radiology and surgery, 11(4), 505-519. https://www.custusx.org

[6] Tomviz. https//www.tomviz.org. [Online; accessed 28-November-2018]

Acknowledgement

This work was funded by Talente (FEMtech Praktika, project no. 869267) via the Austrian Research Promotion Agency (FFG) and supported by the Correlated Multi Modal Imaging Node Austria. VRVis is funded by BMVIT, BMDW, Styria, SFG and Vienna Business Agency in the scope of COMET - Competence Centers for Excellent Technologies (854174) which is managed by FFG.

Visualization of Tumor Vascularization
Figure 1: Visualization result showing anatomical context from µCT, tumor hull, tumor blood vessels from ultra-high field MRI (orange), µCT (red) and HREM (yellow). The additional vessel parts are gained from µCt by manual labeling (green) and thresholding (blue).
Flowchart of the proposed Pipeline
Figure 2: Overview of the processing pipeline consisting of Image Acquisition, Vessel Segmentation, Correlation and Visualization. The orange rectangles depict the imaging modalites, the light blue rectangles show the used software tools and the dark blue circles illustrate the performed processings.
Keywords: tumor vascularization, visualization, image correlation, multi-modal multi-scale imaging, image processing
215

Effect of different ROI definitions on the quantification of SUV (#197)

Bashair Alhummiany1, Richa Gandhi1, Stephen J. Archibald2, Christopher Cawthorne3, Marc A. Bailey1, Charalampos Tsoumpas1

1 University of Leeds, Department of Biomedical Imaging Science, Leeds Institute of Cardiovascular and Metabolic Medicine, Leeds, United Kingdom
2 University of Hull, PET Research Centre, School of Life Sciences, Hull, United Kingdom
3 KU Leuven, Leuven, Belgium

Introduction

Although the standardised uptake value (SUV) has been widely used to quantify tracer uptake in positron emission tomography-computed tomography (PET-CT) images, it can be influenced by the method of defining regions of interest (ROIs)1. Since the maximum SUV can be affected by noise, another method was introduced to improve the accuracy of SUV by applying a threshold to estimate the average SUV of a group of voxels2. Our aim was thus to identify a reproducible threshold value using two segmentation methods. We also assessed the applicability of using these methods in preclinical PET-CT.

Methods

Fourteen mice (10 with abdominal aortic aneurysms, 4 controls) underwent 90-min dynamic PET-CT acquisitions using the Sedecal Super Argus PET-CT scanner. Images were reconstructed with the 3D ordered subset expectation maximisation algorithm (2 iterations, 16 subsets). Using Amide software, all data were segmented using manual and fixed-size ROIs at 80–90 min to compute the mean uptake (SUVmean); maximum uptake (SUVmax); and mean uptake for voxels higher than 40%, 50%, 70%, and 90% of the maximum (SUV40, SUV50, SUV70, and SUV90, respectively) (Fig 1). A fixed 4-mm sphere was used for delineation; the size was chosen based on the maximum aortic aneurysmal diameter (3 mm). Volumetric differences at each threshold and the delineation times were analysed to assess the segmentation methods.

Results/Discussion

By comparing threshold values for each pair of SUVs calculated using the manual and fixed-size methods, significant differences were found at the 40%, 50%, and 90% thresholds (all P < 0.05). However, SUV70showed no statistically significant difference between the two segmentation techniques (P=0.72), indicating that applying a 70% threshold on ROIs can minimise operator bias and result in a more stable SUV measurement. The average volumes of fixed-size ROIs were relatively larger than those of the manual ROIs, since the size of the fixed-size ROIs was chosen to segment different aneurysm diameters. The result also demonstrated that the volumes of ROIs using both methods decreased when a higher threshold value was applied,indicating that fewer voxels were accounted for with a higher threshold.Finally, we found that the average time required for delineation using manual segmentation was up to three times more than that needed to draw the fixed-size ROIs.

Conclusions

In this study, we assessed the reproducibility of four different threshold values for voxels greater than the maximum when two different segmentation methods were used. We concluded that the implementation of fixed-size ROIs with a 70% threshold appears to provide the most stable SUV measurement in the analysis of aortae in preclinical PET-CT images of mice.

References

  1. Boellaard R. "Standards for PET image acquisition and quantitative data analysis." J Nucl Med 50(S1), 2009
  2. Lee JR, et al. A threshold method to improve standardized uptake value reproducibility. Nucl Med Comm 21(7) 2000

 

Placement of ROIs

Figure1 The placement of manual (A and C), and fixed-size (B and D) ROIs on transverse, coronal, and sagittal views in PET-CT images of two different mice with AAA. The bottom panels show AAA that is close to the urinary bladder (UB).

Keywords: PET-CT, Standardised uptake value (SUV), Image segmentation