presentation stage graphe de connectivité du cerveau 2014 (fr)

Click here to load reader

Post on 21-Jun-2015

127 views

Category:

Health & Medicine

0 download

Embed Size (px)

TRANSCRIPT

  • 1. Classification de graphesde connectivit du cerveau Romain Chionencadr par: S. Achard, M. Desvignes, F. Forbes

2. gipsa-labSOMMAIREPRESENTATION DU CONTEXTEMETHODES USUELLESNOUVELLE METHODECOMPARAISON DES RESULTATS 3. CONTEXTEMETHODES3gipsa-labCONTEXTE Comment comparer lesgraphes entre eux? Est il possible de modliserles graphes de connectivitcrbrale (GCC)? A quel point peut-oncatgoriser les GCC? 4. CONTEXTEMETHODES4gipsa-labMODELES GENERATIFSIllustration Small World , Collective dynamics ofsmall-world networks, D. J. Watts & S. H. StrogatzIllustration Preferential Attachement , Choice-driven phasetransition in complex networks, P. L. Krapivsky and S. Redner Erdos-Renyi Forest Fire Kronecker Preferential Attachment Random k-regular Random Power Law Random Typing Small-World 5. COMPARAISON DE GRAPHES Tansformation dun graphe vers un autreex : Distance dditionCONTEXTEMETHODES5gipsa-labMESURESSTRUCTURELLES Tendance des noeuds se regrouper, distributiondes degrs, chemins entre noeudsex : Clustering, Plus Court CheminMESURESLOCALES(pour chaque noeud) Mesures locales moyennes, formation de noyaux et decommunautsex : Assortativit, Centralit, Modularit,DiamtreMESURESGLOBALES 6. ETAT DE LART : JANSSEN et al. 2012Comptage de GraphletsNombre deGraphletsapprentissage du classifieur entre du classifieurMETHODESAPPORTS6gipsa-labEnsembledapprentissageInstance degrapheNombre deGraphletsClassifieurModle de graphe 7. METHODESAPPORTS7gipsa-labETAT DE LART : MOTALLEBI et al. 2013Classifieurde Rseaux Complexes 8. Intervalle deconfiance ~25%METHODESAPPORTS8gipsa-labMODELISATION DES GCCCaractrisation des GCC vers 4 modles(Erdos-Renyi, Preferential Attachement, Random k-regular, Small-World)Classe Prdiction E-R P A R k-R S-WControl Small-World 0.2502 0.2501 0.2492 0.2505Patient Small-World 0.2502 0.2501 0.2492 0.2505 Rsultat de la caractrisation avec mesures globales et classifieur SVM 9. Prcision de la mthode 50.16%, alatoire 50%METHODESAPPORTS9gipsa-labIDENTIFICATION DES GCCtrue Control true Patient class precisionpred. Control 13 11 54.17%pred. Patient 7 6 46.15%class recall 65.00% 35.29% 50.16% Rsultat de lidentification avec mesures globales et classifieur SVM 10. METHODESAPPORTSgipsa-labPROBLEMATIQUE Les mesures globales ne sont pasreprsentatives du comportement local Histogrammes ducoefficient de clusteringlocal pour 3 modles10 11. HISTOGRAMME NORMALISE Clustering Coefficient Characteristic Path Length Degrees Distribution EfficiencyAPPORTSRESULTATS11gipsa-labEnsembleHistogramme desmesures localesdapprentissageInstance derseauHistogrammesnormaliss moyensDistances entreHistogrammesModle de grapheHistogrammenormalisminimum des distances ou un classifieur 12. DISTANCE ENTRE HISTOGRAMMES Mesure de (dis)semblance bin bin : Battacharyya: Chi Hellinger : Mesure de dissemblance avec conservation de la silhouette : EarthMoverDistance : Optimisation du travail minimum qu'un cantonnierdoit fournir pour transporter un tas de terre en un autre Match : Comparaison des histogrammes cumulsAPPORTSRESULTATS12gipsa-lab 13. PerformancesRESULTATSgipsa-labDONNEES DE SYNTHESES13 graphlets : 78% mesures globales : 88% 97.3% 6 mesures voire plus mesures locales : 86% ou 100% 1 seule mesurePrecisionSW 100%RPL 100%RkR 100%PA 100%KG 100%FF 100%ER 100%100%PrecisionSW 100%RTG 96%RPL 98%PA 99%KG 96%FF 98%ER 93%97.2%Rsultat de laclassificationhistogrammes mesures globales 14. GRAPHES DE CONNECTIVITESmesures globales 63% V.S. 83%MAX histogrammesRESULTATS14gipsa-labGLOBALESA.N.N.C PC 11 9 55%P 5 12 71%69% 57% 63%HISTOGRAMMECLUSTERINGET CHIC PC 18 2 90%P 4 13 76%82% 87% 83% Matrice de confusion de lidentification Control / Patient 15. RESULTATSgipsa-labMODELISATION DES GCC157 Clustering DegrsER 0,418 0,133FF 0,207 0,074KG 0,112 0,211RPL 0,156 0,088PA 0,437 0,242RkR 0,459 0,183SW 0,103 0,238 Distance EMD entre GCC et modles pour deux mesures locales 16. Erdos-RenyiFFRPLRESULTATSgipsa-labCLASSE MANQUANTE16ForestFireRPLSWKroneckerGraphFF77%SW23%RPLPreferentialAttachmentFFRPLRandomk-RegularFFRPLRandomPower LawFF92% SW8% PASmall-WorldFFRPLGraphes deConnectivitsFFRPLPASW 17. RESULTATSgipsa-labROBUSTESSE N ET D17100 200 300 400 500 600 700 800 900 100011001200130014001500160017001800190020000,01 11% 10% 10% 14% 14% 12% 7% 6% 14% 11% 23% 29% 29% 30% 29% 34% 34% 36% 36% 45%0,02 12% 18% 18% 16% 20% 22% 30% 39% 41% 42% 43% 42% 42% 44% 42% 43% 43% 42% 42% 43%0,03 10% 19% 20% 27% 28% 41% 41% 45% 43% 43% 43% 42% 41% 40% 44% 43% 44% 43% 43% 43%0,04 17% 26% 32% 40% 43% 41% 44% 40% 43% 43% 43% 43% 43% 43% 45% 43% 43% 43% 43% 42%0,05 16% 25% 41% 42% 41% 43% 42% 43% 38% 40% 43% 42% 42% 43% 42% 43% 42% 43% 43% 43%0,06 33% 41% 43% 44% 43% 42% 42% 46% 41% 43% 43% 43% 42% 43% 43% 43% 49% 43% 44% 43%0,07 36% 57% 54% 65% 62% 70% 67% 72% 71% 72% 69% 71% 68% 72% 85% 85% 83% 86% 84% 86%0,08 44% 69% 72% 72% 72% 75% 69% 86% 84% 86% 86% 86% 86% 86% 86% 86% 84% 86% 86% 86%0,09 41% 81% 85% 93% 96% 93% 90% 97% 94% 90% 86% 86% 84% 85% 71% 71% 70% 72% 71% 71%0,1 49% 88% 86% 100% 96% 100% 99% 84% 81% 85% 86% 86% 86% 86% 86% 86% 86% 86% 86% 86%0,11 52% 99% 93% 90% 89% 91% 92% 78% 74% 72% 71% 71% 69% 71% 71% 71% 71% 71% 71% 71%0,12 62% 83% 85% 72% 71% 68% 72% 68% 74% 73% 72% 71% 71% 71% 72% 71% 71% 72% 71% 71%0,13 62% 64% 70% 64% 68% 68% 71% 68% 67% 67% 70% 66% 69% 57% 65% 61% 56% 43% 48% 43%0,14 59% 57% 48% 43% 43% 44% 44% 44% 43% 43% 43% 43% 43% 43% 43% 43% 43% 43% 43% 43%0,15 54% 49% 45% 49% 42% 42% 45% 42% 42% 43% 43% 43% 43% 43% 43% 43% 43% 43% 43% 43%0,16 45% 44% 43% 43% 44% 45% 42% 44% 43% 43% 43% 43% 43% 43% 43% 43% 43% 42% 43% 43%0,17 42% 41% 41% 40% 42% 42% 42% 42% 43% 44% 43% 43% 42% 41% 41% 42% 42% 41% 39% 39%0,18 45% 43% 44% 43% 43% 45% 42% 42% 43% 42% 43% 41% 41% 37% 40% 37% 34% 32% 31% 31%0,19 44% 45% 43% 39% 43% 42% 41% 42% 42% 40% 36% 33% 30% 32% 29% 29% 29% 29% 29% 29%0,2 43% 43% 41% 45% 41% 41% 40% 35% 30% 35% 31% 29% 29% 29% 29% 29% 29% 29% 29% 29%nombre de noeuds croissantdensit croissante 18. ROBUSTESSE APPRENTISSAGERESULTATS18gipsa-labCROSS-VALIDATION en d100 200 300 400 500 600 700 800 900 1000110012001300140015001600170018001900200067% 70% 67% 69% 70% 71% 73% 75% 74% 77% 77% 78% 77% 78% 79% 80% 80% 78% 79% 81%nombre de noeuds croissantCROSS-VALIDATION en nd = 0,01 0,02 0,03 0,04 0,05 0,06 0,07 0,08 0,09 0,1 0,11 0,12 0,13 0,14 0,15 0,16 0,17 0,18 0,19 0,2PREC 76% 82% 95% 97% 97% 97% 99% 99% 99% 99% 97% 99% 98% 99% 99% 99% 99% 99% 99% 99%MINPREC. 32% 31% 80% 88% 89% 91% 96% 96% 96% 96% 91% 93% 92% 94% 97% 96% 96% 98% 98% 98%MINCLASS. ER ER FF FF KG KG KG SW SW SW KG SW SW SW SW SW SW SW SW SWdensit croissante 19. RESULTATSgipsa-labRANDOMISATION190% 5% 10%15%20%25%30%35%40%45%50%55%60%65%70%75%80%85%90%ER 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100%FF 100% 100% 97% 97% 100% 97% 100% 100% 100% 100% 100% 97% 100% 100% 100% 100% 100% 100% 100%KG 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100%PA 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100%RkR 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100%RPL 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100% 100%SW 100% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%randomisation croissante 20. PC 1 0.415 0.750 0.750PC 2 0.170 0.126 0.876PC 3 0.132 0.076 0.952PC 4 0.101 0.044 0.996PC 5 0.028 0.004 0.999PC 6 0.011 0.000 1.000PC 7 0.003 0.000 1.000RESULTATSgipsa-labPCA : RESULTATS20NBR DE COMPOSANTES PRINCIPALESVARIANCE CUMULEESTRD DEV %VAR VAR 21. AVANT APRESCROSS-VALIDATION d : 5 16% daugmentationgipsa-labmoyenne grimpe de 75 84%RESULTATSPCA : ROBUSTESSE2114% 14% 14% 2% 0% 20% 11% 5% 14% 15% 14% 15% 15% 15% 15% 17% 17% 20% 23%1% 14% 25% 15% 20% 28% 25% 24% 19% 36% 36% 37% 37% 40% 40% 55% 43% 42% 43%6% 26% 31% 30% 35% 43% 46% 63% 62% 64% 63% 58% 60% 61% 67% 61% 66% 63% 61%27% 34% 42% 45% 53% 55% 60% 66% 67% 67% 66% 68% 66% 64% 63% 51% 55% 57% 55%31% 43% 48% 57% 59% 60% 63% 70% 66% 69% 71% 70% 70% 71% 71% 58% 70% 78% 70%32% 51% 56% 69% 70% 66% 68% 72% 71% 74% 72% 86% 86% 85% 84% 71% 85% 86% 83%34% 62% 68% 71% 70% 71% 83% 86% 87% 86% 86% 85% 84% 85% 86% 79% 84% 86% 86%36% 67% 67% 79% 85% 86% 86% 86% 84% 86% 86% 86% 86% 86% 86% 86% 85% 86% 86%40% 76% 94% 99% 100% 100% 98% 99% 99% 100% 100% 100% 99% 100% 94% 96% 96% 83% 82%46% 96% 99% 100% 98% 100% 99% 98% 98% 98% 100% 100% 100% 100% 88% 88% 87% 88% 86%52% 100% 96% 100% 99% 100% 100% 95% 94% 92% 93% 96% 92% 91% 86% 86% 86% 86% 86%57% 98% 100% 99% 100% 98% 87% 73% 74% 77% 75% 72% 72% 73% 72% 71% 71% 72% 71%58% 80% 85% 68% 73% 67% 70% 57% 55% 59% 57% 57% 57% 58% 58% 57% 57% 57% 57%61% 64% 69% 64% 66% 61% 63% 59% 58% 58% 57% 57% 58% 57% 57% 57% 57% 58% 57%65% 57% 67% 62% 59% 60% 58% 56% 58% 57% 57% 58% 57% 57% 58% 58% 58% 58% 57%68% 59% 61% 53% 56% 57% 57% 58% 58% 57% 57% 57% 57% 57% 57% 57% 58% 57% 57%66% 56% 56% 42% 45% 52% 57% 57% 57% 58% 57% 58% 57% 57% 57% 57% 57% 57% 57%62% 57% 61% 43% 43% 47% 54% 58% 58% 55% 57% 57% 57% 57% 58% 58% 58% 57% 57%60% 58% 57% 43% 43% 43% 44% 49% 54% 47% 46% 56% 57% 57% 57% 57% 57% 56% 57%57% 59% 52% 46% 43% 42% 43% 42% 41% 43% 44% 43% 43% 43% 43% 43% 43% 43% 43%11% 10% 10% 14% 14% 12% 7% 6% 14% 11% 23% 29% 29% 30% 29% 34% 34% 36% 36%12% 18% 18% 16% 20% 22% 30% 39% 41% 42% 43% 42% 42% 44% 42% 43% 43% 42% 42%10% 19% 20% 27% 28% 41% 41% 45% 43% 43% 43% 42% 41% 40% 44% 43% 44% 43% 43%17% 26% 32% 40% 43% 41% 44% 40% 43% 43% 43% 43% 43% 43% 45% 43% 43% 43% 43%16% 25% 41% 42% 41% 43% 42% 43% 38% 40% 43% 42% 42% 43% 42% 43% 42% 43% 43%33% 41% 43% 44% 43% 42% 42% 46% 41% 43% 43% 43% 42% 43% 43% 43% 49% 43% 44%36% 57% 54% 65% 62% 70% 67% 72% 71% 72% 69% 71% 68% 72% 85% 85% 83% 86% 84%44% 69% 72% 72% 72% 75% 69% 86% 84% 86% 86% 86% 86% 86% 86% 86% 84% 86% 86%41% 81% 85% 93% 96% 93% 90% 97% 94% 90% 86% 86% 84% 85% 71% 71% 70% 72% 71%49% 88% 86% 100