1 National Institute of Health and Medical Research (INSERM),Multi-parametric multi-modality imaging for prognostic and predictive modeling in oncology Dimitris Visvikis Director of Research National Institute of Health and Medical Research (INSERM), LaTIM, UMR 1101 Brest, France
2 PET/CT multimodality imagingIntroduction Cancer Oncology Gold standard for diagnosis Other applications of interest: Radiotherapy planning Prognosis, therapy assessment PET/CT multimodality imaging Quantification uptake measurement functional tumor volume uptake distribution analysis 2
3 Activity concentration indicesSemi-quantitative index of concentration of activity in a given ROI: ROI choise Robustness Reproducibility Repeatibility No ROI Robustness Reproducibility Repeatibility ROI fixed Robustness Reproducibility Repeatibility
4 Acquisition protocolsDynamic acquisitions: Analysis of dynamic sequence of images Mathematical modelling Arterial sampling Kinetic behaviour of FDG in a particular ROI Input function Parameters adjustment
5 Other popular PET image derived indicesTumour / background Metabolically active tumor volume (MATV, cm3) Total lesion glycolysis (TLG = MATV × SUVmean) SUVmax SUVpeak MATV SUVmean TLG
6 Functional volume segmentationImage noise (differences in acquisition protocols) Partial Volume effects (spatial resolution) Voxel size (4-5 mm) Tumour heterogeneity
7 Thresholding approachesNestle U et al., J Nucl Med, 2005
8 FLAB: Fuzzy Localy Adaptive BayesianProbability of observation P(Y|X) [ , ] (μ,σ) tumor Physiological background Fuzzy transitions Spatial correlation probability P(X) 1 2 3 4 5 6 7 8 9 Final map Hatt M et al. IEEE TMI 2009; Int J Rad Onc Biol Phys 2010
9 Accurate MATV: prediction, prognosisPredictive and prognostic value of functional volume for different cancer models Initial FDG PET scan Better patient stratification and management (survival, therapy response) Example: 45 esophageal cancer patient Therapy response Survival Hatt et al EJNM 2010
10 Accurate MATV: radiotherapyPTVFLAB PTVTHRES Lungs Heart Spinal Cord Le Maitre A et al., Phys Med Biol, 2012
11 Tumor characterisation: geometric formsHypotheses: associated with tumor aggresivity, metastasis potential… morphological, functional and/or morpho-functional: Form descriptors (a)sphericity, solidity, convexity, rectangularity, excentricity… Intra-tumor PET activity distribution form : independent prognostic value demonstrated on H&N cancer1 Lung cancer2 Sarcoma3 CT PET - FDG Fusion PET/CT Apostolova I, et al. BMC Cancer. 2014 Hofheinz F, et al. Eur J Nucl Med Mol Imaging. 2015 Eary J, et al. J Nucl Med 2008
12 Tumor characterisation: geometric forms18FLT PET during chemo-radiotherapy1 High sphericity and low SUV Low sphericity et high SUV Time (months) 12 24 36 48 60 72 Survival probability (%) Predictive value of sphericity2 Hazard ratio = 6,7 (p<0,0001) (with SUV only HR = 4,1, p=0,02) (with sphericity only HR = 4,2, p=0,01) Hoeben BA, et al. J Nucl Med. 2013 Majdoub M, et al. (submitted) 2016
13 Tumor heterogeneity characterisationActivity conc, volume, … Global mesures Intra-tumor activity distribution heterogeneity characterisation
14 Tumor heterogeneity characterisationTextural features can quantify different types of voxels intensity variability in the tumor volume, at different scales First-order parameters describe global textural features that relate to the grey level frequency distribution within the region of interest. They are based on histogram analysis and include mean, minimum and maximum intensity, standard deviation, skewness and kurtosis. Second-order features describe local texture features and are calculated using spatial grey level dependence (SGLDM) or cooccurrence matrices. These matrices determine how often a pixel of intensity i finds itself within a certain relationship to another pixel of intensity j. Second-order features based on a co-occurrence matrices include entropy, energy, contrast,homogeneity, dissimilarity and correlation. Higher-order parameters can be calculated using neighbourhood greytone (intensity) difference matrices (NGTDMs) to describe local features [26, 38]. Local textural parameters derived from NGTDMs are based on differences between each voxel and the neighbouring voxels in adjacent image planes, and are thought to closely resemble the human experience of the image [38]. For example, coarseness, one of the local textural parameters, has been likened to granularity within an image and is the most fundamental property of texture. Contrast relates to the dynamic range of intensity levels in an image and the level of local intensity variation and busyness relates to the rate of intensity change within an image [24, 38]. Regional features can also be calculated from voxel alignment (e.g. run length and run-length variability) and grey level size-zone matrices that reflect regional intensity variations or the distribution of homogeneous regions (e.g. zone emphasis and size-zone variability) [26] (Table 1). M. Tixier F et al, J Nucl Med, 2011
15 Tumor heterogeneity characterisationHistogram analysis No spatial information Versatility and potential Complexity Co-occurrence Matrices Local spatial information Regional spatial information Jensen. Introductory Image Processing 3rd ed. Upper Saddle River, NJ: Prentice-Hall 2005
16 PET/CT image derived featuresParameters for tumor heterogeneity characterisation Textural features can quantify different types of voxels intensity variability in the tumor volume, at different scales. First-order parameters: a = b = c = d Second-order features: a # (b = c = d). Third-order features: a # b # c # d First-order parameters describe global textural features that relate to the grey level frequency distribution within the region of interest. They are based on histogram analysis and include mean, minimum and maximum intensity, standard deviation, skewness and kurtosis. Second-order features describe local texture features and are calculated using spatial grey level dependence (SGLDM) or cooccurrence matrices. These matrices determine how often a pixel of intensity i finds itself within a certain relationship to another pixel of intensity j. Second-order features based on a co-occurrence matrices include entropy, energy, contrast,homogeneity, dissimilarity and correlation. Higher-order parameters can be calculated using neighbourhood greytone (intensity) difference matrices (NGTDMs) to describe local features [26, 38]. Local textural parameters derived from NGTDMs are based on differences between each voxel and the neighbouring voxels in adjacent image planes, and are thought to closely resemble the human experience of the image [38]. For example, coarseness, one of the local textural parameters, has been likened to granularity within an image and is the most fundamental property of texture. Contrast relates to the dynamic range of intensity levels in an image and the level of local intensity variation and busyness relates to the rate of intensity change within an image [24, 38]. Regional features can also be calculated from voxel alignment (e.g. run length and run-length variability) and grey level size-zone matrices that reflect regional intensity variations or the distribution of homogeneous regions (e.g. zone emphasis and size-zone variability) [26] (Table 1). Chicklore, et al. Quantifying tumour heterogeneity in 18F-FDG PET/CT imaging by texture analysis. Eur J Nucl Med Mol Imaging 2013
17 Heterogeneity quantification1. Tumor delineation [1-3] 2. Heterogeneity quantification using textural features [4-6] Co-occurrence matrix For the PET analysis we have used the basline scan done at the diagnosis before any treatment. On these scan we have in a fist time identified the tumors and then delineated we a previously validated algorithm. After that for the quantization we have constructed two matrices: The co-occurrence matrices that describe relationship between contiguous voxels The size-zone matrices that describe relationship between homogeneous areas into the tumor and theirs size 2. Quantization Size-zone matrix [1] Hatt, et al. IEEE Trans Med Imaging. 2009;28: [2] Hatt, et al. Int J Radiat Oncol Biol Phys. 2010;77(1):301-8 [3] Hatt, et al. Eur J Nucl Mol Imaging. 2011;38: [4] Tixier, et al. JNM. 2011;52: [5] Tixier, et al. JNM. 2012;53: [6] Hatt, et al. EJNM. 2013;40:
18 Tumor heterogeneity characterisationM. Tixier F et al, J Nucl Med, 2011
19 Tumor heterogeneity characterisationExample : NSCLC NSCLC without metastasis (n=100) Stade I (n=18), II (n=29), III n=(53) Males (n=78), Females (n=22) Age 64±9 Treatment : Surgery (n=47) Chemotherapy (n=82) Radiotherapy (n=51) Surgery 18 19 1 9 Chemo Radio 12 41
20 Tumor heterogeneity characterisationGlobal survival Médian 18 months Survival probabiity (%) 3 year survival ~ 30% 12 24 36 48 60 Time (months)
21 Tumor heterogeneity characterisationStade I : médiane - p = 0.008 Survival probability (%) Stade II : médian 21 months Stade III : médian 18 months 12 24 36 48 60 Time (months)
22 Multimodality tumor characterisationComplimentarity: functional volume – PET heterogeneity N=34, 17 deaths (50%) Median 45 months N=38, 25 deaths (66%) Median 21 months, HR=2.3 N=29, 27 deaths (93%) Median 9 mo, HR=3.8 Probabilité de survie (%) Time (months) 4 3 2 Entropy 1 Volume PET (cm3) logarithmic scale M. Hatt, et al. J Nuc Med 2015
23 PET/CT image derived featuresHeterogeneity characterization What about CT or other modalities (MRI)? Several studies have investigated the use of textural features or other metrics to characterize CT or MRI tumor volumes1 Colon cancer Contrast CT 1. Davnall, et al. Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice? Insights Imaging 2013
24 PET/CT image derived featuresHeterogeneity characterization What about CT or other modalities (MRI)? Rectal cancer T2-weighted MRI 1. Davnall, et al. Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice? Insights Imaging 2013
25 PET/CT image derived featuresHeterogeneity characterization What about CT or other modalities (MRI)? Esophageal cancer - CT Pre-treatment (baseline) Post-treatment (neoadjuvant chemo) 1. Davnall, et al. Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice? Insights Imaging 2013
26 Multi-parametric & Multi-modalityPrognostic model PET/CT (116 NSCLC patients) Stage, PET volume and heterogeneity, CT heterogeneity Desseroit et al Eur J Nucl Med 2016
27 Therapy follow-up, adaptivePerspectives Future disease characterization with imaging Full tumor « signature » in the multi modal spectrum Multi tracers Multi modalities PET / CT PET / MRI Temporal data Therapy follow-up, adaptive Tracer kinetics Metabolism, hypoxia, proliferation MRI T2, diffusion, elastography…
28 Perspectives Actual Paradigm Proposed Paradigm FusionEx. Astronomy images (3 observations) Proposed Paradigm
29 Perspectives Future disease characterizationCombination and correlation of multimodal tumor signatures with other information Information fusion → Building predictive models
30 Conclusions PET can provide In the future:useful information for early response assessment extract more information from PET requires standardization, image enhancement accurate, robust, reproducible automated delineation new features such as texture or shape In the future: Multi-center trials and large patient cohorts combination of omics data with multimodal anatomo-functional characterization (tumor “signature”) → efficient predictive models for improved patient management
31 Thank you for your attention
32
33 PET tumor heterogeneity: reproducibility
34 Local entropy (co-occurrence matrix)PET tumor heterogeneity: minimum volume Corrélation = 0,56 Corrélation = 0.98 Corrélation = 0.93 Quantization = 64 13 matrices + mean Quantization = 128 13 matrices + mean Quantization = 64 1 matrice Local entropy (co-occurrence matrix) Volume (log, cm3) M. Hatt, et al. J Nuc Med 2015
35 PET tumor heterogeneityMethod Textural features analysis Validation Reproducibility, robustness vs. reconstruction, partial volume effects, tumor delineation [6-8] Application Prediction of therapy response/patient’s outcome from baseline scan[1-5] What does PET heterogeneity represent? Hypothesis: associated with tumor physiology (glucose metabolism, hypoxia, angiogenesis, …) Since few years I was demonstrated that new indices that quantified the heterogeneity on PET images provided an useful information for the patient management These parameters extracted on baseline scan can be use for the prediction of the therapy response or patient’s outcome [1] Tixier F et al. J Nucl Med. 2011; 52: [5] Huang B et al. AJR Am J Roentgenol. 2012; 199: [2] George J et al. ISBI 2012 [6] Galavis PE et al. Acta Oncologica. 2010; 49: [3] Cook GJR et al. J Nucl Med. 2013; 54: 1-8 [7] Tixier F et al. J Nucl Med. 2012; 53: [4] Willaime JMY et al. Phys. Med. Biol. 2013; 58: [8] Hatt M et al. Eur J Nucl Med and Mol Imaging ahead of print