Task 4: Multimodal image analysis

Local tumor control

Multi modal (PET, CT, MRI, …) and multi sequential (various radiotracers) images provide a plethora of information for radiotherapy treatment initial planning [1]. A specific focus will be the use of multi modal images in order to take into account localized tumor characteristics and changes in anatomy and function, not only before but also during treatment. One of the main issues is finding biomarkers able to provide quantitative information from tumor characteristics for improved targeting and assessing the response to the therapy. This novel approach will allow the investigation of the potential use of multi modality imaging in adaptive radiotherapy as a power Aided Decision tool. The goal is to allow the correction of the initial treatment plan according to the information provided by additional images acquired during the treatment [2].

Objective

The objective of this study is to describe, model and quantify morphological changes, the response of the prostate cancer and functional MRI prostatic performed before and during radiotherapy prostatic and hormonotherapy.

 Data acquisition, Treatment protocol

2 groups of patients:
- RT exclusive for intermediate risk tumours (10 patients): 5 MRI will be performed by patient (pre-RT,week 2, week 4, week 6 and week 8), 2 PET (pre-RT, w6), PSAradiotracers prelevements (pre-RT, week 2, week 4, week 6 and week8)
- Hormonotherapy and RT for high risk tumours (10 patients): 2 MRI performed (pre-treatment before hormonotherapy and 2 months after), 2 PET (pre-treatment before hormonotherapy and m2), PSA radiotracers prelevements (pre-treatment before hormonotherapy and m2).
At this stage, 19 patients were included in the protocol (10 RT and 9Hormonotherapy).
 
 

Data processing

Current works could be divided into different step: delineation, registration, feature extraction, feature selection and feature evolution. This work include the extraction of parameters from acquired images and identify those enabling a better differentiation between tissues and a better risk of recurrence prediction.
 
The discriminatory power of each parameter will be evaluated via a study on classification.  Semiautomatic segmentation methods of the prostate and the tumor will also be developed. Project workflow
 
 

Delineation: prostate and tumour location

A single physician segmented each MRI-T2 sequence separately. The structures delineated were:  peripheral Peripheral Zone (PZ),  Transitional Zone (TZ) and prostate tumour (Tumour). Those structures were delineated over the MRI-T2 modality using DWI, ADC and DCE-MRI images to support the diagnosis. Typically,  PCa tumors appear as hypo-intense regions compared to adjacent prostate tissue in MRI-T2,high signal intensity on DWI images and low signal intensity on ADC apparent diffusion coefficient map and had an early enhancement and an early wash-out DCE-MRI.

 

 

Intra-patient image registration

To extract features corresponding to the same regions in each sequence were used two intra-patient registration steps, firstly, the images of DWI, ADC and DCE-MRI were rigid transformed to the MRI-T2space. This inter-modality registration was achieved comparing the mutual information between the images while realizing an Euler transform. Secondly, to analyses only the evolution within the tumour region, the radiologist delineations in MRI-T2 at the initial state were propagated to the other MRI-T2 images corresponding to the same patient. This step of registration used the delineations previously drawn by the physician, by computing a deformation field within the prostate and after performing a rigid registrations.

Feature Extraction

In case DCE-MRI modality, we extract four different features related physical properties of the tissue, namely,wash in coefficient (Ktrans), intra-cellular space (Kev), time necessary to reach the 95% of the maximum value (time to peak) and the total amount of gadolinium uptaken (area under the gadolinium curve) using an in-house implementation of the Toft's model [3]. Extraction of features in other modalities were computed using Halarick's texture features [4].
Current works include the extraction of parameters from acquired images and identify those enabling a better differentiation between tissues and a better risk of recurrence prediction. The parameters discriminatory power will evaluated via a study on classification obtained with different parameters.

Feature Analysis

Each feature was z-normalized w.r.t. the entire voxel population of every patient. Different combination of features of modalities were evaluated, namely: MRI-T2, ADC, DCE-MRI, {MRI-T2, ADC}, {MRI-T2, DCE-MRI}, {ADC,DCE-MRI} and {MRI-T2, ADC, DCE-MRI}. We used
the minimun Redundancy Maximun Relevance (mRMR) to extract meaningful characteristics that represent TZ, PZ and tumour tissue for MRI-T2, ADC, DCE-MRI, {MRI-T2, ADC}, {MRI-T2, DCE-MRI }, { ADC, DCE-MRI} and {MRI-T2, ADC, DCE-MRI}. mRMR uses a heuristic rule to select a short set of features that best discriminates among different regions (K) [5]. This methodology maximizes the mutual information between a set of selected features
(Sj ) and the tissue to which it belongs (maximum relevance O) and minimizes the mutual
information among the features selected (minimum redundancy R).

Segmentation

The arrange of features fed two different classifiers, one to discriminate between TZ and tumour, and another to discriminate between PZ and tumour. The classifiers applied were K-nearest neighbours, linear discriminant analysis and support vector machine, adding one feature in order   of relevance suggested by the mRMR. Finally, we choose the best combination set-of-features /classifier that yields the highest performance.
 
 

Currently we are writing a paper respect to this subject.

Feature evolution during radiotherapy 

Herein, we developed two different approaches to describe the tissue changes due to RT. The first approach consist in use the delineations made by the radiologist on the pre-treatment mpMRI set of images and its corresponding intermodality/intra-patient propagation, were computed the mean of the intensity within pre-treatment images and the following samples, comparing separately TZ, PZ and Tumour tissue. It allows measure the intensity changes due to RT. The second approach used the most significant features to classify between TZ, PZ and tumour regions in pre-therapy images to discriminate same regions in the following samples, allowing us to analyse voxel tumour changes related to other tissues.

 

A second approach consist in relate the most relevant modality to separate tumour voxels from healthy tissue to the PSA marker during RT. In the literature shows that PSA value in blood is related with the tumour and prostate sizes [6]. Herein, we extend this notion by adding the effect of extracellular space and tissue tortuosity measured by the ADC modality, i.e., supposing not uniform cell distribution across the prostate. Those tissue characteristics change due to RT and patient specific responses.
 
 

Currently we are writing a paper about this subject.

 

Conclusion

This study aims to identify MRI indicators based on cumulated radiation dose in prostate. We  also correlated these MRI parameters with biochemical recurrence (PSA serum).
 

 Bibliography

[1] J.L. Lagrange, et al. Image guided radiationtherapy (IGRT). Bull Cancer 2010.
[2] G. Cazoulat, et al.Fromimage-guided radiotherapy to dose-guided radiotherapy. CancerRadiother 2011;15(8);691-8.
[3] Tofts, P., Modeling tracer kinetics in dynamic gd-dtpa mr imaging, Journal of Magnetic Resonace Imaging 7, 91 – 101 (1997).
[4] Haralick, et al, Textural features for image classification. IEEE Trans-actions on systems man and cybernetics 3, 610 – 621 (1973).
[5] Hanchuan Peng, et al. Feature selection based on mutual information: Criteria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on pattern analysi and machine intelligence, 27:1226 – 1238, 2005.
[6] Kristin R. Swanson, et al. A quantitative model for the dynamics of serum prostate-specific antigen as a marker for cancerous growth. An explanation for a medical anomaly. American Journal of Pathology, 158(6):2195 – 2199, June 2001.

Publications and events

Publications

M. Hatt, M. Majdoub, M. Vallières, F. Tixier, C. Cheze Le Rest, D.Groheux, E.Hindié, A. Martineau, O. Pradier, R. Hustinx, R. Perdrisot, R.Guillevin, I. El Naqa, D. Visvikis. FDG PET uptake characterization through texture analysis: investigating the complementary nature of heterogeneity and functional tumor volume in a multi-cancer site patient cohort. J Nucl Med 2015 56(1):38-44.
M-C. Desseroit, D. Visvikis, F. Tixier, M. Majdoub, R. Perdrisot, R. Guillevin, O. Pradier, M. Hatt, C. Cheze Le Rest. Complementary prognostic value of intra-tumor heterogeneity in Non-Small Cell Lung Cancer assessed by combined textural analysis of functional and morphological data using 18F-FDG PET/CT images. Radiology (submitted)
D.Groheux, M. Majdoub, A. Martineau, D. Visvikis, M. Espié, M.Hatt, E. Hindié. Early metabolic response to neoadjuvant treatment: 18FDG-PET/CT criteria according to breast cancer subtype. Radiology 2015 (in press).
D.Groheux, M. Majdoub, A. Martineau, D. Visvikis, M. Espié, M.Hatt, E. Hindié. 18FDG uptake and total lesion glycolysis measured at baseline and after 2 courses of neoadjuvant chemotherapy are powerful tools to predict relapse in patients with ER+/HER2- breast cancer. J Nucl Med (in minor revision)
D. Groheux, M. Majdoub, F. Tixier, C. Cheze Le Rest, A. Martineau, P. Merlet, M. Espié, A. de Roquancourt, E. Hindié, M. Hatt. Do clinical, histological or immunohistochemical primary tumor characteristics translate into different 18FDG-PET/CT image features in stage II-III breast cancer? J Nucl Med (in revision)
Majdoub M, Visvikis D, Tixier F, B. Hoeben, E. Visser, Cheze-Le Rest C, Hatt M. Prognostic value of head and neck tumor shape features derived from 18F-FLT PET/CT images. J Nucl Med (submitted)
Fargeas A, Albera L, Kachenoura A, Dréan G,Ospina J-D, Coloigner J, Lafond C, Delobel J-B, De Crevoisier R,Acosta O. On feature extraction andclassification in prostate cancer radiotherapy using tensordecompositions. Medical Engineering and Physics, 2014 (under review).
Coloigner J, Fargeas A, Kachenoura A, Wang Lu,Dréan G, Lafond C, Senhadji L, De Crevoisier R, Acosta O, Albera L. A novel classification method for prediction of rectal bleeding in prostate cancer radiotherapy based on a semi-nonnegative ICA of 3D planned dose distributions. Journal of Biomedical and HealthInformatics, 2014 (accepted for publication).
Communications in (inter)national peer-reviewed conferences
R. E. Gutiérrez-Carvajal, A. Fargeas, K. Gnep , Y. Rolland, O. Acosta, R. De Crevoiser. A Comparison of Multiparametric MRI Modalities to Discriminate Prostate Cancer Tumours. 10th International Symposium on Medical Information Processing and Analysis 2014
M.C.Desseroit, C.Cheze-Le Rest, F.Tixier, M. Majdoub, R. Guillevin, R. Perdrisot, D. Visvikis, M.Hatt, Complementary Prognostic Value of CT and 18F-FDG PET Non-Small Cell Lung Cancer Tumor Heterogeneity Features Quantified Trough Texture Analysis, American Association of Physicists in Medicine annual meeting- Science Council session, 2014
M.Hatt, F. Tixier, M. Majdoub, C. Cheze-Le Rest, O. Pradier, R. Hustinx, D. Visvikis, Quantification of intra-tumor 18F-FDG PET tracer uptake heterogeneity through texturalfeatures analysis: which minimum functional tumor volume to consider?, Society of nuclear medicine and molecular imaging annual meeting, 2014
M. Majdoub, D. Visvikis, F. Tixier, B. Hoeben, E. Visser, C. Cheze-LeRest, M. Hatt. Proliferative 18F-FLT PET tumor volumes characterization for prediction of locoregional recurrence and disease-free survival in head and neck cancer. Society of nuclear medicine and molecular imaging annual meeting, 2013.
A. Fargeas, A. Kachenoura, O. Acosta, L. Albera, G. Dréan, R. De Crevoisier. Feature extraction and classification forrectal bleeding in prostate cancer radiotherapy: a PCA based method. Recherche en Imagerie et Technologies pour la Santé 2013.
A. Fargeas, L. Albera, A. Kachenoura, O. Acosta, G. Dréan, A. Simon, R. De Crevoisier. Prediction of rectal bleeding in prostate cancer radiotherapy: blind source separation approaches. European Society for Therapeutic Radialogy and Oncology annual meeting 2013.
Fargeas A, Kachenoura A, Acosta O, Albera L, Dréan G, De Crevoisier. Feature extraction and classification forrectal bleeding in prostate cancer radiotherapy: a PCA based method. IRBM, 2013. 34(4–5):p. 296-299.
 

Events

28 Mars 2014 - During a meeting about MRI projects, preliminar results about multimodal MRI analysis (task 4) were presented by Oscar Acosta and Renaud de Crevoisier. The presentation is availiable here (https://servez-vous.univ-rennes1.fr/6hny18am).