Presentation on theme: "Giorgio De Nunzio 1, Marina Donativi 1, Gabriella Pastore 2, Matteo Rucco 3, Antonella Castellano 4, Andrea Falini 4 1.Dept of Materials Science, Univ."— Presentation transcript:
Giorgio De Nunzio 1, Marina Donativi 1, Gabriella Pastore 2, Matteo Rucco 3, Antonella Castellano 4, Andrea Falini 4 1.Dept of Materials Science, Univ. of Salento, and INFN (National Institute of Nuclear Physics) (Lecce, Italy) 2.PO 'Vito Fazzi' - UOC Fisica Sanitaria, and Dept of Materials Science, University of Salento (Lecce, Italy) 3.School of Science and Technologies, University of Camerino (Italy) 4.Neuroradiology Unit and CERMAC, San Raffaele Scientific Institute (Milano, Italy) MAGIC5 meeting - Genova 2011/5/4-6 A CAD system for cerebral glioma and therapy follow-up in DTI and FLAIR: status report
2 Glioma Common primary brain tumors. Typical infiltrative growth pattern glioma cells preferentially infiltrate along white matter fibers. Conventional MRI cannot accurately localize microscopic glioma infiltrations, therefore it does not always permit precise delineation of tumor margins or tumor differentiation from edema and/or treatment effects. Diffusion Tensor Imaging (DTI); Isotropic and Anisotropic maps Diffusion of water molecules
3 Materials and methods Glioma in MR-DTI CAD System for glioma segmentation (texture analysis) Glioma structure during follow-up
4 svoi csvoi rsvoi FEATURE SELECTION Training step ANN (learn) IMAGE ACQUISITION ROI CREATION SLIDING WINDOW MAZDA SCATTER PLOT PCA IMAGE ACQUISITION GUI SLIDING WINDOW MAZDA PCA ANN (classification) CAD ROI CREATION Once upon a time…
5 svoi csvoi rsvoi FEATURE SELECTION Training step ANN (learn) IMAGE ACQUISITION ROI CREATION SLIDING WINDOW 69 FEATURES PCA LDA IMAGE ACQUISITION GUI ENTIRE BRAIN SLIDING WINDOW 69 FEATURE PCA LDA ANN (classification) CAD ROI CREATION ALL IN MATLAB APPLY FILTER FEATURE SELECTION NOW!!!
6 Fisher’s score vs feature
7 PCA or LDA? Work in progress…
8 Why Principal Component Analysis? Maximize variance by axis transformation
9 Dimensionality Reduction Can ignore the components of lower significance. You do lose some information, but if the eigenvalues are small, you don’t lose much –n dimensions in original data –calculate n eigenvectors and eigenvalues –choose only the first p eigenvectors, based on their eigenvalues –final data set has only p dimensions Variance Dimensionality
10 Limitations of PCA Are the maximum-variance variables the relevant features for discrimination preservation?
11 Linear Discriminant Analysis What is the goal of LDA? −Perform dimensionality reduction “while preserving as much of the class discriminatory information as possible”. −Seeks to find directions along which the classes are best separated. −Takes into consideration the within-class scatter but also the between-class scatter.
12 PCA given an s-dimensional vector representation (features) of each sample in a training set, Principal Component Analysis (PCA) tries to find a s-dimensional space whose basis vectors correspond to the maximum-variance directions in the original feature space. The dimensionality of this new space is then normally decreased to a lower one (t << s) by neglecting directions with low eigenvalues. If x is the feature array, it is possible to diagonalise the covariance matrix: and obtain the eigenvalues of the linear transformation Matrix T, that is Starting from that, it is possible to calculate the PC’s Feature dimensionality reduction methods LDA Linear Discriminant Analysis finds the vectors in the space that best discriminate among classes. For two classes, the solution proposed by Fisher is to maximize a function that represents the difference between the means, normalized by a measure of the within-class scatter Within-class scatter matrix Between-class scatter matrix
14 PCA LDA
15 … Classifier:Artificial Neural Networks (ANN) AUC=0.94 AUC=0.97 FLAIR 5 patients for training and 4 for test Back-propagation feed-forward ANN: 1 hidden layer, with 3 neurons 1 output neuron PCA LDA
16 … some results of map creation and segmentation Probability maps in p, q or FLAIR images: the dots mark the positions of the sliding window (svoi centers). Color scale: darker colors for low probability values, lighter colors for high values. Red line: shows the segmentation produced by the CAD system (“arbitrary” threshold)
17 P MAP - PCAP MAP - LDA … some results of map segmentation P MAP 6 patients for training and 6 for test AUC=0.88AUC=0.95 P MAP – MED ROI
18 Q MAP - PCAQ MAP - LDAQ MAP – MED ROI … some results of map segmentation Q MAP 6 patients for training and 6 for test AUC=0.77 AUC=0.90
19 FLAIR - PCAFLAIR - LDAFLAIR – MED ROI … some results of map segmentation FLAIR 5 patients for training and 4 for test AUC=0.94AUC=0.97 Fluid Attenuated Inversion Recovery
20 Prospects Feature reduction or selection? PCA Fisher score LDA ICA Both selection and reduction?? Fisher score (with a threshold according to the AUC plateau [as a function of the FS]) Jaccard Coefficient to set the ‘best’ ANN threshold for segmentation FLAIR segmentation is promising!FLAIR segmentation is promising!
21 Changes in glioma water diffusion values after chemotherapy: work in progress!! LGG (low-grade glioma) cells grow and diffuse typically along the white matter tracts Diffusion Tensor Imaging in glial tumors allows to depict white matter alterations not visible by conventional MRI Price et al., Clin Radiol 2003; Wang et al., AJNR 2009 Starting from Diffusion Tensor it is possible to obtain two maps: isotropic (p) and anisotropic (q) Isotropic (p) and Anisotropic (q) Maps allow a better characterization of the diffusion features of tumoral and peritumoral areas Pena et al., BJR 2006; Price et al., Eur Radiol 2004 Price et al., AJNR 2006; Price et al., Eur Radiol 2007; Wang et al., AJNR 2009
22 Changes in tumor water diffusion occur after successful treatment and can be attributed to changes in cell density. increase of tissue infiltration decrease of tissue infiltration p q p q Moffat et al., PNAS 2005 Hamstra et al., JCO 2008 Galban et al., TransOnc 2009 Aim of Study: to investigate whether changes in the Brownian motion of water within tumor tissue as quantified by using diffusion MRI could be used in the follow up of treated gliomas. Changes in glioma water diffusion values after chemotherapy: work in progress!!
23 IDAgeSite Previous surgery Histology 1p- 19q MGMTSeizures TMZ (dose- dense) 1 34Frontal LOct-05O IIcodelN/ANo6 cycles 2 28Frontal LJul-07OA IIcodelmetNo6 cycles 3 33 Fronto-tempo- insular R Jul-04O II N/A No6 cycles 4 25 Fronto-tempo- insular L Sep-07O II no codel N/A Yes6 cycles 5 36Frontal LApr-07A II no codel metYes6 cycles 6 56 Fronto-tempo- insular L Jul-04A II N/A Yes6 cycles 7 45Frontal LNov-04O II N/A Yes6 cycles 8 37 Fronto-tempo- insular L Sep-09OA II no codel unmetYes6 cycles 9 32 Fronto- temporal R Mar-08OA II no codel N/ANo6 cycles 9 patients with low grade glioma similar Histology same duration of treatment similar Clinical History same scheme of neuroradiological follow-up Patients & methods
24 Patients & methods 3T Scanner Intera Philips Medical System (gradients 80 mT/m) MR morphological study: axial T2 TSE (TR/TE 3000/85, NSA 2), axial FLAIR (TR/TE/TI /120/2800) axial FFE MP-RAGE (TR/TE 8/3.9) voxel size and positioning as for DTI, acquired following i.v. injection of paramagnetic contrast DTI scans: axial Single-Shot Spin Echo EPI (TR/TE 8986/80, b-value 1000 mm 2 /sec, 32 directions, SENSE 2.5, FOV 240, mm, repeated twice) Diffusion maps: diffusion-tensor elements calculated and diagonalized at each voxel, obtaining three eigenvalues, fractional anisotropy (FA), and trace (Tr) maps; from the elaboration of these datasets in MATLAB pure isotropic (p) and pure anisotropic (q) diffusion maps are obtained Segmentation of tumor areas in the various maps of tensor decomposition metrics (p, q) obtained from first MR examination and after five cycles IMAGE ACQUISITION IMAGE COREGISTRATION SEARCH OF THRESHOLD PIXEL COLOR MAP q and p MAP TUMOR AREA SEGMENTATION
25 y>x p value before chemoterapy x y x y y
26 ID Seizure response Radiological response (FLAIR) DTIDTI response % blue/red voxels on isotropy map Second surgery extent of resection (%) Peritumoral IDH1 1N/ASD+ of isotropy 7.3 blue; 6.1 red 79,72 subtotalN/A 2 mR -36.4%+ of isotropy 2.7 blue; 2.2 red 83,48 subtotalN/A 3 mR -26%- of isotropy 26 blue; 53.5 red 82,67 subtotalN/A 4StableSD -11.6%+ of isotropy 27 blue;13 red82,16 subtotalN/A 5StablePD- of isotropy 1 blue; 30 red97,5 subtotalN/A 6 >50% SD -9,7%+ of isotropy 5.8 blue; 3.7 red 97,7 subtotalN/A 7 >50% mR -35%+ of isotropy 4.4 blue; 3.1 red 100 totalNeg 8 >50% SD+ of isotropy 7 blue; 4.8 red100 totalNeg 9N/ASD+ of isotropy 67 blue; 17 red79 subtotalN/A Results
27 voxel blu: 7% voxel rossi: 4.8 voxel blu: 7% voxel rossi: 4.8% M.G., astrocitoma WHO II: fDM su mappa p (isotropia) dopo 6 cicli TMZ I esame II esame stabilità radiologica di malattia miglioramento clinico
28 voxel blu: 1% voxel rossi: 30 voxel blu: 1% voxel rossi: 30% R.G., oligodendroglioma WHO II: fDM su mappa p (isotropia) dopo 6 cicli TMZ progressione di malattia! I esame II esame progressione radiologica di malattia? stabilità clinica
29 voxel blu: 7.3% voxel rossi: 6.1 voxel blu: 7.3% voxel rossi: 6.1% I esame II esame B.L., oligodendroglioma WHO II: fDM su mappa p (isotropia) dopo 6 cicli TMZ stabilità radiologica di malattia miglioramento clinico
30 sonda monopolare per la stimolazione intraoperatoria infiltrazione fibre CST B.L., ODG WHO II: confronto con neurofisiologia intraoperatoria infiltrazione fibre CST sonda bipolare per la stimolazione intraoperatoria
31 Tissue analysis with DTI could help the physicians in evaluating the chemotherapy responses.Tissue analysis with DTI could help the physicians in evaluating the chemotherapy responses. The p or q value variation could suggest a tumorThe p or q value variation could suggest a tumor progression or regression also for cases in which the progression or regression also for cases in which the tumor volume does not change. tumor volume does not change. These preliminary results are in accordance withThese preliminary results are in accordance with neurophysiological results and with intraoperative neurophysiological results and with intraoperative bioptic samples. bioptic samples. Some conclusions Some conclusions :
32 Prospects: we are working to… unify in the scatter plot both the p and the q value variations study also the local maximum variation
33 Conferences: 1.A. CASTELLANO, M. DONATIVI, L. BELLO, G. DE NUNZIO, M. RIVA, G. PASTORE, G. CASACELI, R. RUDÀ, R. SOFFIETTI, AND A. FALINI(2011). Evaluation of changes in gliomas structural features after chemotherapy using DTI-based Functional Diffusion Maps (fDMs): a preliminary study with intraoperative correlation. In:2011 Joint Annual Meeting ISMRM-ESMRMB. Montréal, May 7-13, G. DE NUNZIO, M. DONATIVI, G. PASTORE, A. CASTELLANO, A. FALINI, L. BELLO, R. SOFFIETTI, (2010). A CAD system for cerebral glioma and therapy follow-up in Diffusion-Tensor Images. In II Workshop Plasmi Sorgenti Biofisica e Applicazioni, Lecce (Italy) 26 Ottobre A. CASTELLANO, L. BELLO, E. FAVA, G. CASACELI, M. RIVA, M. DONATIVI, G. PASTORE, G. DE NUNZIO, R. RUDA', L. BERTERO, R. SOFFIETTI, A. FALINI (2010) DTI-MR 3D Texture Analysis per la valutazione delle modificazioni delle caratteristiche strutturali dei gliomi cerebrali dopo trattamento con Temodal: studio preliminare. In XV Congresso Nazionale della Associazione Italiana di Neuro-Oncologia (AINO), Fiuggi (FR, Italy) 3-6 Ottobre G. DE NUNZIO, G. PASTORE, M. DONATIVI, A. FALINI, A. CASTELLANO, L. BELLO, R. SOFFIETTI (2010). DT-MR images: A CAD System for Cerebral Glioma and Therapy Follow-up. In IVth European Conference of Medical Physics - Advances in High Field Magnetic Resonance Imaging, Udine (Italy) September (2010) 5.G. DE NUNZIO, M. DONATIVI, G. PASTORE, A. CASTELLANO, G. SCOTTI, L. BELLO, A. FALINI (2010). Automatic Segmentation and Therapy Follow-up of Cerebral Glioma in Diffusion-Tensor Images. In 2010 IEEE International Conference on Computational Intelligence for Measurement Systems and Applications (CIMSA 2010). Taranto (Italy) September 6-8, CASTELLANO, L. BELLO, E. FAVA, M. RIVA, G. CASACELI, G. DE NUNZIO, M. DONATIVI, G. PASTORE, R. RUDA', R. SOFFIETTI, A. FALINI (2010) Changes in gliomas structural features after Temodal treatment evaluated by DTI-MR texture analysis: a preliminary study. In 9th International Meeting UPDATES IN NEURO-ONCOLOGY, Brain Tumor Symposium, Cortona (AR, Italy), July 2-4, G. DE NUNZIO, G. PASTORE, M. DONATIVI, A. CASTELLANO, A. FALINI (2010). A CAD system for cerebral glioma based on texture features in DT-MR images. In International Conference on Imaging Techniques in Subatomic Physics, Astrophysics, Medicine and Biology (Imaging 2010), Stockholm (Sweden) 8-11 June G. DE NUNZIO, A. CASTELLANO, G. PASTORE, M. DONATIVI, G. SCOTTI, L. BELLO, A. FALINI (2010). Semi-automated evaluation of structural characteristics and extension of cerebral gliomas using DTI-MR 3D Texture Analysis. In: 2010 Joint Annual Meeting ISMRM-ESMRMB. Stockholm, May 1-7, G. DE NUNZIO, A. CASTELLANO, M. DONATIVI, G. PASTORE, A. FALINI. (2010). A semi-automated DTI-based approach to evaluate structural characteristics and extension of cerebral gliomas (poster No C-2926). In: European Congress of Radiology (ECR2010). Vienna, March 4-8, G. DE NUNZIO, G. PASTORE, A. CASTELLANO, M. DONATIVI, A. FALINI (2010). Automatic Segmentation of Cerebral Glioma in DT-MR Images by 3D Texture Analysis. In: Risonanza magnetica in medicina: dalla ricerca tecnologica avanzata alla pratica clinica (Italian Chapter of the International Society of Magnetic Resonance in Medicine). Milano, 4-5 febbraio 2010
35 Metodi per la riduzione dello spazio delle feature ai fini della classificazione ICA decompone il dataset nelle sue sottoparti indipendenti Dato il vettore x, mistura dei segnali originali s tramite una matrice di mixing A scopo della ICA è identificare una matrice di de-mixing W tale che le componenti del vettore in uscita siano quanto più statisticamente indipendenti
36 ICA Per stimare una delle IC La combinazione lineare delle sorgenti indipendenti è più “gaussiana” delle componenti originarie e lo diventa “al minimo” quando z ha solo l’i-imo elemento non nullo: questo porta a scegliere W in modo da massimizzare la non-gaussianità di W T x
38 Linear Discriminant Analysis Within-class scatter matrix Between-class scatter matrix projection matrix −LDA computes a transformation that maximizes the between-class scatter while minimizing the within-class scatter: products of eigenvalues ! : scatter matrices of the y data after projection
39 Linear Discriminant Analysis −Since S b has at most rank C-1, the max number of eigenvectors with non-zero eigenvalues is C-1 (i.e., max dimensionality of sub- space is C-1) Does S w -1 always exist? −If S w is non-singular, we can obtain a conventional eigenvalue problem by writing: −In practice, S w is often singular since the data are image vectors with large dimensionality while the size of the data set is much smaller (M << N )
40 Features in MaZda VS Features in Matlab: an example