15th European Molecular Imaging Meeting
supported by:

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

Data Processing & Quantification

Session chair: Bartosz Leszczyńskn (Krakow, Poland); Martin Meier (Hannover, Germany)
Shortcut: PS 16
Date: Thursday, 27 August, 2020, 10:00 a.m. - 11:30 a.m.
Session type: Parallel Session


Abstract/Video opens by clicking at the talk title.

10:00 a.m. PS 16-01

Introductory Lecture

Markus Aswendt1

1 University Hospital Cologne, Cologne, Germany

10:18 a.m. PS 16-02

Assessment of within-breath pulmonary acinar deformation by dynamic in vivo synchrotron lung microscopy in anesthetized rat

Jose Luis Cercos-Pita1, 2, Luca Fardin1, 2, Anders Larsson1, Gaetano Perchiazzi1, Alberto Bravin2, Sam Bayat3, 4

1 Uppsala University, Hedenstierna Laboratory, Department of Surgical Sciences, Uppsala, Sweden
2 European Synchrotron Radiation Facility, Medical Beamline (ID17), Grenoble, France
3 Inserm UA7 STROBE Laboratory, Grenoble, France
4 Grenoble University Hospital, Grenoble, France


In vivo micromechanics of individual lung acini remain largely unknown, mainly due to the lack of microscopic imaging techniques allowing for sufficiently high temporal and spatial resolution. We previously developed a time-resolved synchrotron radiation X-ray phase-contrast tomographic technique which allows to image the lungs in vivo down to the alveolar length scale. Here we describe image processing methodology allowing for the computation of acinar displacement and strain during ventilation, along with initial results in rat lung.


The experiment was performed in anesthetized, muscle-relaxed and mechanically ventilated adult rats. X-ray projection images were acquired at a constant frame rate using a fast PCO camera connected with a visible light optics. 4-D images were reconstructed with 6 µm pixel-size space resolution and 10 ms time resolution. An image analysis pipeline was developed in Python, to segment terminal bronchi and subtending acinar structures at a reference time instant and carry out incremental registration-based voxel-by-voxel displacement computation. The pipeline was also designed to compile such results and compute the deformation of the acinar structures along time.


Figure 1 shows a composite image of 3-D renderings of three terminal airways and subtending acini, with a 2-D raw image slice in greyscale. Figure 2 shows a quantitative map of the regional acinar volume change between the minimal and maximal pressure points during the ventilation cycle.


Our results demonstrate for the first time, the feasibility of mapping the regional acinar volume change in intact in vivo rat lungs at microscopic spatial resolution. This methodology will help elucidate alveolar micromechanics in both healthy lungs and pathologic conditions.

AcknowledgmentThe research leading to these results has received funding from the Swedish Reseach Council under grant 2018-02438 "Mikrodynamiska acinära förändringar i in vivo modeller av lungsjukdomar, studerade med synkrotron stråle datortomografi, positron emissions tomografi och magnetkamera".
Figure 1

Surface rendering of segmented terminal airways and subtending acini overlayed on a grey-level CT image. Blue and light grey: conducting airways; brown: Alveolar ducts and acini.

Figure 2
Reltaive volume change between end-expiration of the segmented airspaces from end-expiration (5 cmH2O) to peak pressure on inspiration (11 cmH2O), overlayed on an axial grey-scale CT image of the lung. Color indicates relative volume change.
Keywords: Micromechanics, Lung, Strain, Acini, Mechanical Ventilation
10:30 a.m. PS 16-03

An advanced machine learning algorithm for the automated detection of tissue chromophores from Multi-spectral photoacoustic images

Valeria Grasso1, 2, Joost Holthof1, Jithin Jose1

1 FUJIFILM Visualsonics, Amsterdam, Netherlands
2 Hannover Medical School, Institute for Animal Science, Hannover, Germany


To detect the tissue chromophores, multi-spectral Photoacoustic (PA) imaging frequently uses a differential based unmixing methods with a known spectral signature as a-priori information1. For the translational research with human patients, these types of supervised spectral unmixing can be challenging, as the spectral signature of the tissues differs with respect to disease condition. So here we present a machine learning approach, the non-negative matrix factorization (NNMF), for the automated detection of tissue chromophores. This is the first time NNMF is using to explore the PA images.


The multi observations (M) for the NNMF are modeled as: M=AS where A is the weight matrix and S is the spectra matrix, which are iteratively updated under the non-negativity constraints2. In this study, the feature extraction capability of other unsupervised algorithms3,4 such as PCA, ICA, has been tested on multi spectral PA images and compared with the NNMF approach. Multi-spectral PA images were acquired by using VevoLAZR-X (FUJIFILM Visualsonics, Toronto). A tissue mimicking phantom containing tubes with different dyes (ICG and Methylene Blue) was used to evaluate the sensitivity of the algorithms. In-vivo experiments were also performed by injecting Indocyanine green (ICG) intravenously and the NNMF has been used to unmix and quantify the endogenous and exogenous tissue chromophores.


The different unsupervised machine learning approaches have been tested by performing the tissue mimicking phantom imaging. Spectral images where obtained with in the wavelength range of 680-900 nm with a step size of 5 nm. PCA, ICA and NNMF were applied to unmix the contrast agents and the Signal to noise of each algorithm was evaluated. According to the SNR values, the performance of the NNMF approach was prominent. Furthermore, the kidney-spleen region of the mouse was imaged. Single wavelength PA images were obtained at 890 nm before and after the injection of ICG. From the images it’s clear that the contrast agent is accumulating in the region as the PA signal intensity is increased. NNMF algorithm was applied on the spectral images and Fig.1 shows the spectrally unmixed components after the injection of ICG. The prominent tissue chromophores like oxy and deoxy hemoglobin are clearly visible in the spectra together with the signature of ICG and it is visualized in the Fig. 2.


Relying on the non-negative nature of the PA images, the constraint of the NNMF unsupervised machine learning approach has ensured high feature detection performance. Here, for the first time, NNMF algorithm was tested on multi-spectral PA images and the automated detection of tissue chromophores was performed. In contrast to other algorithms, NNMF offers superior sensitivity to unmix and it is promising to translate into the clinical measures.


This work has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 811226.

[1] Luke, G.P., Nam, S.Y. and Emelianov, S.Y., 2013. Optical wavelength selection for improved spectroscopic photoacoustic imaging. Photoacoustics, 1(2), pp.36-42.
[2] Lee, D.D. and Seung, H.S., 1999. Learning the parts of objects by non-negative matrix factorization. Nature401(6755), p.788.
[3] Montcuquet, A.S., Herve, L., Garcia, F.P.N.Y., Dinten, J.M. and Mars, J.I., 2010. Nonnegative matrix factorization: a blind spectra separation method for in vivo fluorescent optical imaging. Journal of biomedical optics15(5), p.056009.
[4] Glatz, J., Deliolanis, N.C., Buehler, A., Razansky, D. and Ntziachristos, V., 2011. Blind source unmixing in multi-spectral optoacoustic tomography. Optics express19(4), pp.3175-3184.
Figure 1
Figure1: (A) Expected spectral curves of oxy-deoxy hemoglobin and ICG; (B) Unmixed spectral curves obtained by NNMF
Figure 2

Figure 2: (A) Single wavelength PA image at 890nm; (B) Spectrally unmixed oxygenated hemoglobin; (C) deoxygenated hemoglobin; (D) ICG

Keywords: Photoacoustic imaging, Data quantification, Image processing
10:42 a.m. PS 16-04

PET and MRI evaluation of experimental stroke: a new SPM analysis

Luca Presotto1, Sara Belloli1, 2, Alba Grayston3, Paolo Rainone1, 4, Valentino Bettinardi1, Luigi Gianolli1, Marco Bacigaluppi5, Antonella Castellano6, Nicoletta Anzalone6, Anna Rosell3, Maria Picchio1, 7, Rosa Maria Moresco1, 2, 8

1 IRCCS San Raffaele Scientific Institute, Nuclear Medicine Unit, Milano, Italy
2 CNR (National Research Council), Institute of Molecular Bioimaging and Physiology (IBFM), Segrate, Italy
3 Vall d'Hebron Research Institute, Neurovascular Research Laboratory and Neurology Department, Barcelona, Spain
4 University of Milan, Doctorate School of Molecular and Translational Medicine, Milano, Italy
5 IRCCS San Raffaele Scientific Institute, Neuroimmunology Unit, Milano, Italy
6 IRCCS San Raffaele Scientific Institute, Neuroradiology Unit, Milano, Italy
7 Vita-Salute San Raffaele University, Milano, Italy
8 University of Milano-Bicocca, Medicine and Surgery Department and Milan Centre for Neuroscience, Milano, Italy


Ischemic stroke is one of the main causes of mortality and disability with limited therapies available. Advanced imaging techniques such as PET and MRI can be applied to monitor disease course and to guide therapy. In this work, we characterized a transient stroke model in C57BL/6 male mice. Imaging was performed using MRI and PET/CT with [18F]FDG and [18F]VC701 radioligands, to assess glucose metabolism and microglia/macrophages activation, respectively. A statistical parametric mapping (SPM) approach was introduced for the analysis


Ten mice underwent surgery for MCAO using a transient 45 minutes occlusion protocol. Seven animals completed the imaging study following this scheme: [18F]VC701-PET at day 1; T1-w and T2-w MRI and [18F]FDG-PET at day 2; MRI and PET with both tracers at 1 and 2 weeks post-ischemia. Another group of 10 control mice underwent the same imaging procedures. PET images were analyzed twice: with the classical region of interest (ROI) analysis and with the SPM approach. For the first, 1 mm-diameter spherical region are manually drawn in the lesion center, guided by the MR. In SPM, all images are realigned to a standardized anatomical space, intensity scaled to the mean of the whole brain, and a voxel-wise student t-test identifies regions with abnormal uptake at single subject level


CT-based spatial normalization was effective in aligning mice brains. ROI analysis found a decrease in glucose metabolism in the ipsilateral versus contralateral hemisphere at d2 (-62% of [18F]FDG uptake), partially restored thereafter to -32% at d9 and d16 on ROI quantification, (see Fig. 1). Maximum inflammation was observed at d9, with a 50% increase of [18F]VC701 uptake, remaining stable the week after (+34% at d16). Interestingly also SPM analysis showed a large volume of reduced [18F]FDG uptake, with mean peak t-values of 17, 7.8 and 13 at  d2, d9 and d16, respectively. Applied to [18F]VC701 tracer, SPM found no significant uptake at d1, similarly to the ROI analysis, and significant uptake in the ipsilateral hemisphere at d8 and d15 (mean t-value of 8.5 and 9.5). The two PET images analysis are consistent. The introduction of the SPM pipeline allows a more streamlined and standardized analysis method, automatically detecting the regions affected by stroke (e.g. in Fig 2).


SPM single subject analysis provides an optimal tool to automatically estimate damage in ischemia models. This method avoids operator-dependent errors, allows automatic calculation of lesion volume, and could be oh help in therapy evaluation. This paper applied for the first time an automated SPM analysis to preclinical TSPO imaging in Parkinson's Disease

Acknowledgment “MAGBBRIS” Project, funded by the EuroNanoMed 3 ERA-NET call.
Figure 1
PET images at the various time points of a representative animal with the two difference PET tracers. The evolution of the tracer uptake over time can be observed.
Figure 2
Example output of one SPM analyis at one specific timepoint for a single animal
Keywords: semiquantification, tspo, preclinical PET, SPM
10:54 a.m. PS 16-05

Assessment of Different Quantification Methods for [18F]-NaF PET/CT Images of Patients with Abdominal Aortic Aneurysm

Mercy I. Akerele1, Nouf A. Mushari1, Rachael O. Forsythe2, 3, Marc R. Dweck2, 3, Maaz Syed2, 3, David E. Newby3, Charalampos Tsoumpas1

1 University of Leeds, Biomedical Imaging Science Department, Leeds, United Kingdom
2 University of Edinburgh, British Heart Foundation Centre for Cardiovascular Science, Edinburgh, United Kingdom
3 University of Edinburgh, Edinburgh Imaging Facility, Edinburgh, United Kingdom


Spill-in contamination from the active bone into the aneurysm is a major concern in [18F]-NaF PET/CT imaging of abdominal aortic aneurysms (AAA) as this causes inaccurate quantification in the aneurysm. Past studies have shown that common quantification metrics used in clinical investigation are prone to the spill in effects. Therefore, this study aims to correct for the spill-in effect from the bone into the aneurysm and also to assess alternative quantification metrics which will be less prone to the spill in effect in the absence of spill-in correction.


Sixty-five patients diagnosed with AAA were included in this study. The patient data were reconstructed with the ordered subset expectation maximization (OSEM) algorithm incorporating point spread function (PSF) modelling using software for tomographic image reconstruction (STIR). The spill-in effect in the aneurysm was investigated using two target regions of interest (ROIs): one covering the entire aneurysm (AAA), and the other covering the aneurysm but excluding the part close to the bone (AAA-exc). The spill in effect was corrected using the background correction (BC) technique. Quantifications of uncorrected (PSF) and corrected (PSF+BC) images using different threshold target-to-background (TBR) values (TBRmax, TBR90, TBR70, and TBR50) were compared at 3 and 10 iterations.


The result showed a significant difference in quantification between AAA and AAA-exc which is partly due to the spill-in effect from the bone into the aneurysm. TBRmax showed the highest sensitivity to the spill-in effect while TBR50 showed the least. The spill-in effect is reduced by using 10 iterations instead of 3 iterations, but this comes at the expense of reduced contrast and increased noise. However, TBR50 gave the best trade-off between increased contrast-to-noise ratio (CNR) and reduced spill-in effect. By comparing methods, the PSF+BC method clearly reduced TBR sensitivity to the spill-in effect, especially at 3 iterations in comparison to the application of PSF alone (P-value ≤ 0.05) for all the thresholded TBRs. At 10 iterations, no significant difference was observed between the two methods (P-value > 0.05). Also, PSF+BC showed consistent quantification at 3 and 10 iterations than PSF. In all cases, TBR50 was the most robust metric for reduced spill-in and increased CNR.


The spill-in effect can be effectively reduced by increasing the number of iterations, but this comes at the expense of reduced contrast and increased noise. Applying 3 iterations of PSF+BC and using TBR50 for quantification may be able to reduce the TBR sensitivity to the spill-in effect while maintaining low level of noise, hence improving PET quantification.


M. I. Akerele was supported by a PhD scholarship from Schlumberger Foundation Faculty for the Future, The Netherlands. C. Tsoumpas is supported by a Royal Society Industry Fellowship (IF170011). The SoFIA3 study was funded by Chief Scientist Office (CSO; ETM/365). R. O Forsythe, M. R. Dweck, M. Syed and D.E. Newby are supported by the Medical Research Council (11/20/03). British Heart Foundation (FS/14/78/31020, FS/18/31/33676, CH/09/002/26360, RG/16/10/32375, RE/18/5/34216) and the Wellcome Trust (WT103782AIA).

The authors would like to thank Drs Chengjia Wang and David Senyszak (University of Edinburgh, UK) and Dr Karakatsanis (Weil Cornell Medicine, New York), for their help in processing the patient data.

[1] Liu Y. (2012) “Invalidity of SUV measurements of lesions in close proximity to hot sources due to shine-through effect on FDG PET-CT interpretation,” Radiology Research and Practice vol. 2012, no. 868218, pp. 1-4.
[2] Silva-Rodríguez J., Tsoumpas C., Domínguez-Prado I., Pardo-Montero J., Ruibal Á. and Aguiar P. (2016), "Impact and correction of the bladder uptake on 18F-FCH PET quantification: a simulation study using the XCAT2 phantom," Physics in Medicine & Biology, vol. 61, no. 2, pp. 758.
[3] Akerele M. I., Wadhwa P., Silva-Rodriguez J., Hallett W. and Tsoumpas C. (2018), “Validation of the physiological background correction method for the suppression of the spill-in effect near highly radioactive regions in positron emission tomography,” European Journal of Nuclear Medicine and Molecular Imaging Physics, vol. 5, no. 1, pp. 34.
[4] Forsythe R. O.,  Dweck M. R.,  McBride O. M. B., Vesey A. T.,  Semple S. I., Shah A. S. V., et al. (2018), “18F-Sodium Fluoride Uptake in Abdominal Aortic Aneurysms: The SoFIA3 Study,” Journal of the American College of Cardiology,vol. 71, no. 5, pp. 513-523.
[5] Akerele M. I., Karakatsanis N. A., Forsythe R. O., Dweck M. R., Syed M. Aykroyd R. G., et al. (2019), “Iterative reconstruction incorporating background correction improves quantification in [18F]-NaF PET imaging of patients with abdominal aortic aneurysm,” Journal of Nuclear Cardiology, https://doi.org/10.1007/s12350-019-01940-4.
Sample CT and PET images of a patient dataset

CT images and PET reconstructed images of a patient dataset, showing a high [18F]-NaF uptake in the bone and the aneurysm. The activity contribution from the bone was removed in PSF+BC. The ROIs used to extract the SUVs at the aneurysm are shown on the CTAC image. The outer yellow and inner red ROIs represent AAA and AAA-exc, respectively. Following past research, AAA-exc was drawn such that its distance from the bone is approximately 4mm. The blue sphere is the background ROI used for blood bool correction and the calculation of TBR.

Comparison of TBR metrics for the two reconstruction techniques
Comparisons of the different TBR metrics using 2 ROI delineations. (a) and (b) show the PSF at 10 and 3 iterations respectively, while (c) and (d) show the PSF+BC at 10 and 3 iterations respectively. The equation of the fit lines are also presented in each plot for the TBR metrics, showing the slope and the intercept.
Keywords: Spill-in effect, background correction, quantification, Abdominal aortic aneurysm, target-to-background ratio (TBR)
11:06 a.m. PS 16-06

An artificial intelligence approach for the automatic segmentation of human vasculature in clinical multispectral optoacoustic tomography images

Nikolaos Kosmas Chlis1, Angelos Karlas1, Nikolina-Alexia Fasoula1, Michael Kallmayer3, Hans-Henning Eckstein3, Fabian J. Theis2, Vasilis Ntziachristos1, Carsten Marr2

1 Helmholtz Center Munich, Institute of Biological and Medical Imaging, Munich, Germany
2 Helmholtz Center Munich, Institute of Computational Biology, Munich, Germany
3 Rechts der Isar Hospital, Clinic for Vascular and Endovascular Surgery,, Munich, Germany


Multispectral Optoacoustic Tomography (MSOT) is a potent tool for vascular imaging, capable of resolving oxy- and deoxy-hemoglobin in vivo. Nonetheless, MSOT suffers from gradual signal attenuation with increasing depth which hinders the precise manual segmentation of vessels in deeper tissues. Additionally, vascular assessment via functional tests results in thousands of images rendering their manual segmentation an extremely laborious process. Thus, we introduce herein a deep learning approach for automatic vascular segmentation in clinical MSOT images.


We developed and employed a sparse UNET architecture to automatically segment both arteries and veins in acquired MSOT images, while simultaneously identifying important illumination wavelengths. The network transforms each 400x400x28 input image (Height x Width x Illumination Wavelengths) into a 400x400 binary image showing the final segmented region. To facilitate the feature selection process, the first layer consists of a 1x1 2D convolution of one filter and no bias and enforces both L1 regularization and non-negative weights. We train and evaluate the S-UNET on subsets of raw MSOT images which were acquired from volunteers who consented with the work safety regulations of the Helmholtz Center Munich.


The ground truth segmentation masks correspond to blood vessels (arteries and veins) and were acquired from expert clinicians on the MSOT images under co-registered ultrasound guidance. The model’s success was quantified by calculating the Dice score between the true and predicted segmentation masks. An ensemble of 100 S-UNET instances was trained to ensure stability of the selected wavelengths. The S-UNET ensemble achieved a Dice score comparable to the performance of a standard UNET of similar size. Additionally, the benefit of the S-UNET compared to the standard UNET is that it identified the spectra of total blood volume (810nm) and oxy-hemoglobin (850nm) as the most important wavelengths for the segmentation task. This result aids in the interpretability of S-UNET results since both aforementioned wavelengths have known physical meaning in the case of vascular segmentation.


The above results showcase that the proposed S-UNET method has comparable segmentation performance to a typical UNET, with the added benefit of interpretability via the identification of important wavelengths. Both wavelengths identified by the S-UNET as important have physical meaning in the case of vascular segmentation. To the best of our knowledge, this is the first application of deep learning on clinical MSOT data.

Keywords: arftificial intelligence, deep learning, machine learning, multispectral optoacoustic tomography, optoacoustics