EMIM 2018 ControlCenter

Online Program Overview Session: PS-12

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

Data Analysis & Methodology

Session chair: Daniel Razansky - Munich, Germany; Felix Gremse - Aachen, Germany
 
Shortcut: PS-12
Date: Wednesday, 21 March, 2018, 6:15 PM
Room: Lecture Room 4/5 | level -1
Session type: Parallel Session

Abstract

Click on an contribution preview the abstract content.

6:15 PM PS-12-1

Introductory Talk by Amir Rosenthal - Haifa, Israel

This talk provides an overview of state-of-the-art research and refers to the following presentations selected from abstract submissions.

6:45 PM PS-12-2

Eigenvector centrality mapping and seed-based analysis of resting state fMRI during acute brainstem-coma recovery in the rat (#319)

P. Pais1, 4, B. Edlow2, Y. Jiang1, J. Stelzer1, M. Zou3, X. Yu1

1 Max Planck Institute - Cybernetics, High Field MRI, Tuebingen, Germany
2 Massachusetts General Hospital, Neurosciences intensive Care Unit, Boston, Massachusetts, United States of America
3 Wenzhou Medical University, Wengzhou, China
4 Graduate Training Center of Neuroscience, Tuebingen, Germany

Introduction

Despite the known association between brainstem lesions and coma, a circuit-based understanding of coma pathogenesis and mechanisms of recovery is lacking1. We recently developed a model of coma in the rat with focal injury to the brainstem, which allows investigating the neural mechanisms of coma emergence and recovery2. Resting state functional MRI (rs-fMRI) experiments along coma evolution in the rat provided evidence for an acute recovery mechanism by which subcortical arousal centers outside the brainstem reactivate the cerebral cortex.

Methods

rs-fMRI scans were acquired during the first 8 hours post-coma using a 3D EPI sequence (TE, 12.5 ms; TR, 1s; matrix size, 48x48x32; resolution, 400x400x600 µm; 925 TRs) on a 14.1 T/26 cm magnet interfaced to an Avance III console. Pre-processing was performed in AFNI3 and Lipsia4. For each rs-fMRI scan, a voxel-wise map of eigenvector values was computed (eigenvector centrality map), indicating the importance of the respective voxel within the network, followed by least squares fit regression of the eigenvector values over the temporal succession. This resulted in a certain slope, informative of the increase or decrease in connectivity at a given voxel of the brain (Fig.1). Additionally, seed-based analysis was performed by calculating the Pearson's correlation coefficient between regions.

Results/Discussion

The eigenvector centrality mapping-based whole brain functional connectivity analysis showed increases along the acute recovery from coma in thalamus, basal forebrain and basal ganglia (Fig.1). Additionally, seed-based analysis revealed higher correlations along the post-coma period between the central and reticular thalamus, striatum, globus pallidus and the nuclei in the basal forebrain (Fig.2). Interestingly, the time courses of these nuclei increased their correlation with those in cingulate and somatosensory cortex only after 4 hours post-coma (Fig.2). Concurrent electrophysiology and behavioral assessment in the rats demonstrated recovery of the neurological function during the period of study. This result provides evidence for the participation of the thalamic-basal forebrain-basal ganglia network in recovery of consciousness during acute recovery from coma, a time window that is not accessible in the clinical practice for systematic study of the human brain function.

Conclusions

The convergent results from whole brain and seed-based fMRI analysis of connectivity highly suggest a potential role for the basal forebrain-basal ganglia-thalamocortical network in the initial phase of restoration of consciousness after brainstem injury. This study further verifies the applicability of the rat brainstem coma model to investigate brain dynamics during the acute phase of coma.

References

1. Schiff, N. D. 2010; 2. P.Pais et al 2017; 3. Cox, R. W. 1996; 4. Lohmann, G. et al. 2001.

Acknowledgement

Graduate Training Center of Neuroscience Tuebingen.

Figure 1. Analysis of the whole brain rs-fMRI in rats recovering from coma
A: experimental design and general principle of the slope map. B: averaged slope maps of the comatose animals and of a control group anesthetized with 2% isoflurane. C: brain map showing the z-statistic from the 2-tailed 2-sample t-test between coma and control slope maps.
Figure 2. ROI-specific analysis of rs-fMRI
The central graph summarizes the strengthened connections during the acute recovery phase of coma. The seed-based graphs show the connectivity between specific regions at 4 different post-coma times.
Keywords: brain injury, functional connectivity, arousal, consciousness
6:55 PM PS-12-3

Radiomics for the Discrimination of Tuberculosis Lesions (#207)

P. Gordaliza1, 2, J. J. Vaquero1, 2, S. Sharpe3, M. Desco Menéndez1, 2, 4, A. Muñoz-Barrutia1, 2

1 Universidad Carlos III de Madrid, Bioingenieria e Ingenieria Aeroespacial, Leganés, Madrid, Spain
2 Instituto de Investigación Sanitaria Gregorio Marañón, Laboratorio de Imagen Médica, Madrid, Madrid, Spain
3 Public Health England, Microbiology Services Division, Porton Down, England, United Kingdom
4 Centro de Investigaciones Cardiovasculares Carlos III, Madrid, Madrid, Spain

Introduction

Tuberculosis (TB) is an infectious disease with a high incidence and mortality1. Traditionally, TB has been considered a binary disease, latent/active, due to the limited specificity of the traditional diagnostic tests2. Computer Tomography (CT) images of TB infected subjects presents specific manifestations4 and radiomics techniques can be applied for a more discriminant TB characterization5. Here, we proposed a methodology to automatically extract informative features and discriminate between five different types of lesions.

Methods

The main steps of our pipeline are (Fig.1): 1) Automatic lung segmentation and airway tree extraction6; 2) Selection of relevant volumes (i.e., TB lesions) employing Statistical Region Merging7; 3) Extraction of 26 texture features from each volume at 6 grey level quantizations (L=[8,16,32,64,128,256])8; 4) Parameter optimization of the Random Forest (RF) classifiers9 at each L employing different number of features (100-fold cross validation). Inherent data imbalance is handled employing Tomek Links10; 5) Evaluation of the classifiers performance using the F1-score to distinguish among 5 types of lesions: granulomas, conglomerations, trees in bud, consolidations and ground glass opacities. Selection of  informative features employing the Gini Importance (IG) given by each optimal RF.

Results/Discussion

Fig.2.a) shows the IG of each feature at each quantization level for the optimal RF classifier. Features become much informative for large L’s (difference var., L=8,IG=0.14; information measure of correlation 1, L=256,IG= 0.45). Fig.2.b) depicts the weighted F1 obtained in function of the number of features. At large L’s, a few number of features (6) are enough to reach a good precision (L=256,F1=0.844) close to convergence. These results indicate that using the proposed methodology, it would be possible to longitudinally characterize  the disease based on the discrimination among diverse TB manifestations. These lesions would be characterized by the computation of a small number of features which is crucial to build interpretable models.  Namely, the analysis of the feature Info. Measure Correlation 1, particularly high at the largest L’s (L=256,F1=0.45), allows to characterize complex relationships between adjacent voxels that constitute a unique signature of each type of lesion.

Conclusions

In conclusion, the proposed radiomics framework shows encouraging results in its ability to extract informative features to characterize tuberculosis. In particular, we have proved that it is possible to achieve a reasonable good discrimination of the most frequent TB lesion types and more importantly, that our model achieves an effective quantification of the changes that occur in the lung. In the future, this work will be the base for further studies on the characterization of the biological changes induced by TB infection that would lead to an improved understanding of the disease’s course.

References

 

  1. World Health Organization and others, “Global tuberculosis report 2016,” Tech. Rep., 2016.

  2. Barry 3rd, C. E., Boshoff, H., Dartois, V., Dick, T., Ehrt, S., Flynn, J.,Young, D. (2009). The spectrum of latent tuberculosis: rethinking the goals of prophylaxis. Nature Reviews. Microbiology, 7(12), 845–855. http://doi.org/10.1038/nrmicro2236

  3. Pai, M., Behr, M. A., Dowdy, D., Dheda, K., Divangahi, M., Boehme, C. C.,Raviglione, M. (2016). Tuberculosis. Nature Reviews Disease Primers, 2. http://doi.org/dx.doi.org/10.1038/nrdp.2016.76

  4. Nachiappan, A. C., Rahbar, K., Shi, X., Guy, E. S., Mortani Barbosa Jr, E., Shroff, G. S.,Hammer, M. M. (2017). Pulmonary Tuberculosis: Role of Radiology in Diagnosis and Management. RadioGraphics, 37(1), 52–72. http://doi.org/10.1148/rg.2017160032

  5. Gillies, R. J., Kinahan, P. E., & Hricak, H. (2016). Radiomics: Images Are More than Pictures, They Are Data. Radiology, 278(2), 563–577. http://doi.org/10.1148/radiol.2015151169

  6. Artaechevarria, X., Blanco, D., Pérez-Martín, D., De Biurrun, G., Montuenga, L. M., De Torres, J. P., Ortiz-De-Solorzano, C. (2010). Longitudinal study of a mouse model of chronic pulmonary inflammation using breath hold gated micro-CT. European Radiology, 20(11), 2600–2608. http://doi.org/10.1007/s00330-010-1853-0

  7. Nock, R., & Nielsen, F. (2004). Statistical Region Merging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(11), 1452–1458.

  8. Saeys, Y., Abeel, T., & de Peer, Y. (2008). Robust Feature Selection Using Ensemble Feature Selection Techniques. In ECML PKDD (pp. 313–325). http://doi.org/10.1007/978-3-540-87481-2_21

  9. Ma, L., & Fan, S. (2017). CURE-SMOTE algorithm and hybrid algorithm for feature selection and parameter optimization based on random forests. BMC Bioinformatics, 18(1), 169. http://doi.org/10.1186/s12859-017-1578-z

  10. He, H., & Garcia, E. A. (2009). Learning from Imbalanced Data. IEEE Transactions on Knowledge and Data Engineering, 21(9), 1263–1284. http://doi.org/10.1109/TKDE.2008.239

 

Acknowledgement

The research leading to these results received funding from the Innovative Medicines Initiative (www.imi.europa.eu) Joint Undertaking under grant agreement no. 115337, whose resources comprise funding from EU FP7/2007-2013 and EFPIA companies in kind contribution. This work was partially funded by projects TEC2013-48552-C2-1-R, RTC-2015-3772-1, TEC2015-73064-EXP and TEC2016-78052-R from the Spanish Ministry of Economy, Industry and Competitiveness (MEIC), TOPUS S2013/MIT-3024 project from the regional government of Madrid and by the Department of Health, UK. The CNIC is supported by the MEIC and the Pro CNIC Foundation, and is a Severo Ochoa Center of Excellence (SEV-2015-0505).

Figure 1: Fully-automatic radiomics workflow for the extraction of informative features on the lung
Results
Keywords: Tuberculosis, Radiomics, Computer-Tomography, Feature selection, Texture analysis
7:05 PM PS-12-4

Tumor heterogeneity as a PET-biomarker predicts overall survival of pancreatic cancer patients (#173)

E. M. M. Smeets1, J. Teuwen1, J. A. W. M. van der Laak2, M. Gotthardt1, F. Ciompi2, E. H. J. G. Aarntzen1

1 Radboud university medical center, Department of Radiology and Nuclear Medicine, Nijmegen, Netherlands
2 Radboud university medical center, Department of Pathology, Nijmegen, Netherlands

Introduction

Pancreatic ductal adenocarcinoma (PDAC) shows a 5-year survival rate of 8% [1], mostly due to the lack of effective treatment options [2]. PDAC has a remarkable fibrotic reaction [3] that impacts tumor metabolism, which can be measured on PET/CT images [4]. To date, standard PET-derived parameters, e.g. SUVmax, have not been able to provide prognostic information. In this study, we developed a logistic regression model based on FDG-PET texture features (TF) that classifies PDAC as heterogeneous or homogeneous and shows a good correlation with overall survival.

Methods

Patients with histologically proven PDAC (n=121) who underwent 18F-FDG PET/CT (Siemens Biograph mCT, Knoxville, US) were selected from the hospital system. Eighty-six EANM reconstructed scans [5] were visually labeled as ‘homogenous’ (n=40) or ‘heterogeneous’ (n=46) by two experienced Nuclear Medicine physicians.  In order to extract features, tumors were first delineated using 40% threshold of the SUVmax with manual correction. TF previously shown to be robust with respect to PET scan variability [6-8] were extracted using the PyRadiomics toolbox [9]. A logistic regression classifier was build and applied to the 35 cases held out. The training set was classified via leave-one-out cross validation. Prognostic impact was assessed by Kaplan Meier survival analysis and log-rank test.

Results/Discussion

Optimal performance of the leave-one-out cross-validation classifier in the training set yielded an accuracy of 0.73 and AUC of 0.71 in classifying PDAC as heterogeneous or homogeneous tumors. For the 121 patients the overall survival of PDAC tumors classified as heterogeneous was significantly worse than for homogeneous tumors; median overall survival 69 weeks (95%CI 64 to 91 weeks) versus median 95 weeks (95%CI 76 to 114), p= 0.0285). This is in contrast with single standard PET parameters (SUVmax, kurtosis, metabolic tumor volume), single TF (entropy) or visual labeling, which had no significant prognostic impact (see figure 1).

Conclusions

We developed an algorithm that accurately classifies PDAC as metabolic heterogeneous or homogeneous, based on a set of 18F-FDG PET derived texture features. We showed that the classification result has prognostic value, improving upon standard PET derived parameters and single texture-features. Further validation of this algorithm in an external cohort of PDAC patients warranted.

References

  1. Siegel, R.L., K.D. Miller, and A. Jemal, Cancer statistics, 2016. CA Cancer J Clin, 2016. 66(1): p. 7-30.
  2. Ryan, D.P., T.S. Hong, and N. Bardeesy, Pancreatic adenocarcinoma. N Engl J Med, 2014. 371(11): p. 1039-49.
  3. Neesse, A., et al., Stromal biology and therapy in pancreatic cancer: a changing paradigm. Gut, 2015. 64(9): p. 1476-84.
  4. Heid, I., et al., Co-clinical Assessment of Tumor Cellularity in Pancreatic Cancer. Clin Cancer Res, 2017. 23(6): p. 1461-1470.
  5. Boellaard, R., et al., FDG PET and PET/CT: EANM procedure guidelines for tumour PET imaging: version 1.0. Eur J Nucl Med Mol Imaging, 2010. 37(1): p. 181-200.
  6. Yan, J., et al., Impact of Image Reconstruction Settings on Texture Features in 18F-FDG PET. J Nucl Med, 2015. 56(11): p. 1667-73.
  7. Leijenaar, R.T., et al., The effect of SUV discretization in quantitative FDG-PET Radiomics: the need for standardized methodology in tumor texture analysis. Sci Rep, 2015. 5: p. 11075.
  8. Grootjans, W., et al., The Impact of Optimal Respiratory Gating and Image Noise on Evaluation of Intratumor Heterogeneity on 18F-FDG PET Imaging of Lung Cancer. J Nucl Med, 2016. 57(11): p. 1692-169
  9. van Griethuysen, J.J.M., et al., Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res, 2017. 77(21): p. e104-e107.

Acknowledgement

This study was supported by a grant from the Radboud Oncology Fonds/Stichting Bergh in het Zadel (KUN2015-8106).

Tumor heterogeneity classification based on logistic regression has a prognostic value.
Kaplan-Meier analysis for overall survival according to of PET-derived parameters (a) SUVmax, (b) kurtosis, (c) metabolic tumor volume in cm3, (d) texture-feature derived entropy, (e) visual labeling by two Nuclear Medicine physicians, (f) texture-feature derived classifier.
Keywords: Pancreatic cancer, PET/CT, metabolic heterogeneity, texture features, machine learning
7:15 PM PS-12-5

Dynamic PET Data Analysis Without Frames (#185)

J. L. Herraiz1, E. Fernandez-Garcia2, M. A. Morcillo3, J. M. Udías1

1 University Complutense of Madrid, Nuclear Physics Group, Madrid, Spain
2 University of Oviedo, Science Department, Oviedo, Spain
3 Research Centre for Energy, Environment and Technology - Ciemat, Dep. of Environment, Madrid, Spain

Introduction

Dynamic PET studies provide useful information of the evolution of the biodistribution of tracers in the body. Standard dynamic analysis is performed dividing the acquired data into several time frames which are reconstructed independently. The relevant parameters of a region-of-interest (ROI) are obtained fitting the activity concentration in each frame to a specific function. Nevertheless, this approach requires many image reconstructions, and the use of short frames usually produces noisy images with significant bias. In this work, we propose a novel method to address these problems.

Methods

In our method, each event of the list-mode data is weighted based on the time t they were detected, and then histogrammed into standard sinograms. Weights can be chosen to be, for instance, {1, t, t2} to create the zero, first, and second-order momentum sinograms respectively. The zero-order momentum sinogram corresponds to a single-frame. These sinograms are reconstructed normally, and as they contain all the available data they do not suffer from significant noise or bias. The dynamic parameters of interest in a ROI can be then easily derived from the resulting images using simple algebraic relations. The method has been verified with data from many different preclinical and clinical scanners.

Results/Discussion

As a first example, we used our method to determine the initial activity of a decaying 13N cylinder acquired for 5 minutes with the Biograph scanner. We compared these results with the ones obtained using 10 frames of 30 seconds. Both methods yield similar results, but the estimated uncertainty of our method is smaller. As an example of its application in preclinical studies, we performed a PatLak analysis in the myocardium region of a rat injected with 18F-FDG and acquired in the Argus PET/CT scanner. Using the last 10 out of 44 dynamic frames and an image-derived input function provided a value of Ki=0.091 min-1. We obtained the same result (Ki=0.093 min-1) with our method, but using only 2 reconstructed images, corresponding to the zero and first-order momentum of the acquired data.

Conclusions

The proposed method is a completely new approach to dynamic analysis. Instead of reconstructing multiple images and then fit their values to a particular function, we directly reconstruct the images using the weights needed for the fit. It can be applied to many different dynamic studies using the appopriate set of weights. In cases with a well-determined protocol, such as the PatLak analysis described before, it can be a very effective way to reduce the computational cost and bias in the results.

References

Walker MD et al. Bias in iterative reconstruction of low-statistics PET data: benefits of a resolution model. Phys Med Biol. 2011 Feb 21;56(4):931-49.

Patlak Analysis: http://doc.pmod.com/pxmod/pmclass.lib.pmod.models.PMpatlakV2.htm

 

Acknowledgement

This work was supported by Comunidad de Madrid (S2013/MIT-3024 TOPUS-CM), Spanish Ministry of Science and Innovation, Spanish Government (FPA2015-65035-P, RTC-2015-3772-1). This is a contribution for the Moncloa Campus of International Excellence. Grupo de Física Nuclear-UCM, Ref.:  910059. This work acknowledges support by EU's H2020 under MediNet a Networking Activity of ENSAR-2 (grant agreement 654002). J. L. Herraiz was also funded by the EU Cofund Fellowship Marie Curie Actions, 7th Frame Program.

Motivation of the method

The method is inspired in the way least-square fit of scattered data {xi,yi} is performed, in which the relevant variables of interest are not the individual values of each data point but the sum of {xi, yi,xi*yi,xi*xi}. Therefore, list-mode data do not need to be histogrammed into frames to be analyzed.

Description of the method and Patlak Plot
The proposed method incorporate the dynamic information into the sinograms and therefore it can be recovered from the reconstructed images. It eliminate the need to perform multiple image reconstructions. In many cases two or three images is enough. As an example, our equivalent method to the Patlak plot is derived. 
Keywords: Dynamic PET, Frames, Patlak plot, pharmacokinetics
7:25 PM PS-12-6

Fast Magnetic Particle Imaging using Angular Subsampling Based Reconstruction (#332)

R. Orendorff1, E. Frenklak2, E. Y. Yu1, Y. Shi3, B. Zheng1, S. M. Conolly1, 2

1 University of California Berkeley, Bioengineering, Berkeley, California, United States of America
2 University of California Berkeley, Electrical Engineering and Computer Science, Berkeley, California, United States of America
3 Beijing Institute of Technology, School of Information and Electronics, Beijing, China

Introduction

Magnetic Particle Imaging (MPI) is a promising new modality that images only a magnetic tracer, commonly super-paramagnetic iron oxide (SPIO) nanoparticles [1, 2]. This technique requires manipulating a magnetic field gradient over space to acquire projections of the tracer concentrations in the imager (similar to pencil beam CT), taking an hour or more acquire 3D data [3]. We propose a novel and efficient field free line (FFL) MPI 3D image reconstruction method via prox-linear optimization to significant undersample of the required number of projections.

Methods

A simulation of the Berkeley 7 T/m FFL MPI scanner was used to create sample images of a coronary phantom. The required number of projections to satisfy the Nyquist sampling theorem was calculated and used to simulate a image acquisition with an SNR of 50 using 25nm diameter iron particles. Image acquisition with fewer projections was also simulated, down to 25 projections. The acquired data was then reconstructed using the filtered back-projection (FBP) and using the proposed sparse sampling algorithm.

The novel sparse sampling algorithm solves a prox-linear [5] optimization problem that minimizes a data consistency term and a sparse domain (total variation on the image). The algorithm is matrix-free by a in-house package called PyOp [4] that enables low memory use and fast reconstruction.

Results/Discussion

The coronary phantom was reconstructed using FBP and the proposed SS MPI method. 100 projections was calculated as the minimum number of projections satisfying the Nyquist sampling theorem. The SS reconstruction achieves similar results as the fully sampled image without streak artifacts in a fourth of the number of projections. Sinogram data was simulated with a 25nm iron particle (0.95mm FWHM in a 6.3 T/m FFL MPI scanner) with a 4cm by 4cm, 64x64 pixel FOV. 2D reconstruction time is 1 second; a tilted 3D version of the coronary phantom takes ~10 seconds for a 64x64x64 pixel image.

 

The results show that while FBP for images sampled with ~1/4 the number of projections leads to significant artifacts, the SS method correctly reconstructs the coronary phantom, leading to a 4x acquisition time reduction. As each projection on the Berkeley FFL MPI takes approximately 30 seconds to acquire, this reconstruction allows for 3D datasets to be taken in approximately ten minutes.

Conclusions

The SS MPI algorithm enables fast and robust reconstruction of undersampled, noisy FFL MPI data in a manner that reconstructs the underlying data more accurately than the state of the art reconstruction method. The SS method took 10 seconds to reconstruct a 3D version of the phantom, demonstrating that this algorithm is fast in practice. In addition, the novel forward model can be used to reconstruct data taken using any arbitrary scan trajectory, opening up new possibilities such as incrementally/locally updating images and constant SNR trajectories.

References

[1] P.W. Goodwill et al. IEEE Trans. Med. Imaging, 2010

[2] T. Knopp et al. Magnetic Particle Imaging, 2012

[3] J. J. Konkle et al. Biomed Tech (Berl). 2013

[4] R. Orendorff et al. WMIC 2016

[5] A. Beck et al. SIAM J. Imaging Sciences. 2009

Acknowledgement

We are grateful for funding support from the Keck Foundation Grant 009323, NIH 1R01EB019458, NIH 1R24MH106053, and a UC Discovery Grant. Ryan Orendorff would like to thank the NSF GRFP for funding support. 

Model of FFL MPI for Sparse Subsampling Algorithm
Sparse sampling (SS) reconstruction algorithm for FFL MPI. (a) Image of the Berkeley FFL MPI scanner simulated in this study. (b) The FFL MPI forward model is calculated from a generalized Radon transform that allows for arbitrary trajectories of space. The model is used in a prox-linear algorithm to reconstruct the image given some penalty on the TV norm.
Angular Sparse Subsampling (SS) Algorithm faithfully reconstructs coronary phantom

The results of the reconstruction on a simulated phantom image. 100 projections was calculated as the Nyquist rate, given a 25nm MPI tracer. The SS (25 projection) reconstruction achieves similar results as the fully sampled (100 projection) image without streak artifacts, enabling a 4x reduction in image acquisition time.

Keywords: Magnetic Particle Imaging, Sparse Subsampling, Image Reconstruction, Convex Optimization
7:35 PM PS-12-7

Tumor Detection and Characterization in Ultrasound Data by Automated Feature Extraction and Pattern Recognition (#538)

Z. A. Magnuska1, T. Opacic1, F. Kiessling1, B. Theek1

1 Uniklinik RWTH Aachen, Experimental Molecular Imaging, Aachen, North Rhine-Westphalia, Germany

Introduction

The primary goal of radiomics is to improve precision medicine by developing classification models, which are supporting physicians in diagnosis and therapy evaluation. Until now, only a few radiomic analysis has been carried out for ultrasound (US) data [1,2]. Therefore, we developed an advanced, user-independent workflow to conduct a radiomic analysis of US images and evaluated its capability to automatically differentiate three different experimental tumor models.

Methods

US B-mode images of 3 different subcutaneous tumor models (lung cancer(n=3), ovarian cancer(n=3), SCC(n=3)) scanned at 2 positions, were retrospectively analyzed. Cascade classifiers (CCs) were trained for an automated tumor detection (Fig. 1A) [3]. The detection box was the seeding point for the following tumor segmentation, which is based on an active contour model and morphological operations [4]. From the segmented region, a total of 230 intensity-based, textural and wavelet-based features were mined (Fig. 2). Dedicated feature selection was performed to identify three linearly independent features for the final tumor classification model. The generated radiomic signature (RS) was validated with a k-NN learning algorithm, based on the leaving-one-out cross-validation scheme.

Results/Discussion

The proposed algorithm for the automated detection and segmentation of tumors, which was based on the Viola-Jones algorithm and an active contour segmentation, achieved a high detection accuracy (89% of correct tumor detections), and automated segmentations overlapped with the manual segmentations (81% ± 3% overlap) (Fig. 1B). The developed imaging biomarker extraction and selection algorithm identified the RS consisting of the following three independent features: median (intensity-based feature), correlation (textural feature) and short run emphasis (wavelet-based feature). The isolated tumor classification model assigned correctly 80% of tumors to their histological group (p=0.8 [95% CI 0.6-0.9]) (Fig. 2).

Conclusions

Radiomic analysis of US images can be performed to classify tumors, and should be more extensively evaluated for clinical translation. The developed framework might be able to standardise tumor detection and segmentation and to support clinicians in tumor recognition and characterisation.

References

[1] Andrekute K. et al. Ultrasound Med Biol 2016, 42(12): 2834-2843.

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

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

[4] Kass, M et al. Int J Comput Vision 1988, 1: 321-331.

Figure 1
Automated Tumor Detection and Segmentation
Figure 2
Workflow of Automated Tumor Differentiation
Keywords: radiomics, automated segmentation, artificial intelligence, machine learning, image analysis