IronTract Challenge - Round I results

Transcript

1. The IronTract challenge: Validation and optimal tractography methods for the

   HCP diffusion acquisition scheme C. Maffei , G. Girard , K. G. Schilling , N. Adluru , D. B. Aydogan , A. Hamamci , F. Yeh , M. Mancini , Y. Wu , A. Sarica , A. Teillac , S. H. Baete , D. Karimi , Y. Lin , F. Boada , N. Richard , B. Hiba , A. Quattrone , Y. Hong , D. Shen , P. Yap , T. Boshkovski , J. S. W. Campbell , N. Stikov , G. B. Pike , B. B. Bendlin , A. L. Alexander , V. Prabhakaran , A. Anderson , B. A. Landman , E. J.Z. Canales-Rodrígue , M. Barakovic , J. Rafael-Patino , T. Yu , G. Rensonnet , S. Schiavi , A. Daducci , M. Pizzolato , E. Fischi- Gomez , J. Thiran , G. Dai , G. Grisot , N. Lazovski , A. Puente , M. Rowe , I. Sanchez , V. Prchkovska , R. Jones , J. Lehman , S. Haber , A. Yendiki RESULTS FROM ROUND I

2. Kissing Fibres Fanning Fibres Crossing Fibres Courtesy of Hui Wang

   Introduction: Tractography limitations Previous challenges: • Limited angular or spatial resolution • Single acquisition scheme • Comparison using different brains IronTract challenge goals: • Understand which methods/processing strategies lead to optimal tractography accuracy for the two-shell HCP scheme • Investigate whether those methods could achieve even higher accuracy with a different acquisition scheme.

3. Training + Validation Case Anatomical Specificity Anatomical Complexity Methods Consistency

   DSI + Multishell diffusion schemes DSI Grid Multishel l Evaluate wider range of methods Compare different sampling schemes Test methods for HCP sampling scheme True Positive Rate False Positive Rate Tractography results at different thresholds Investigate the Sensitivity-Specificity Tradeoff Compare different methods at the same FPR ROC Analysis Introduction: Why is this challenge unique?

4. HCP Scheme Ex-vivo diffusion MRI NU-IFFT DSI Grid Training Case

   Validation Case Resampled Methods: The challenge timeline 4.7T Bruker 3D EPI 0.7 mm iso DSI HCP 0.0 0.1 0.2 0.3 AUC Score Parameter Optimization Fessler & Sutton ‘03 bmax =40000s/mm2 515 Volumes bmax =6000,12000s/mm2 186 Volumes

5. Registered Teams = 30 Final Submissions = 17 Teams: 12

   Total Submissions = 226 Training = 187 Validation = 39 H D H D Results: Submission Details

6. Results: ROC analysis TrackMcTrackface TwoPathsDverged Accesschallenge X-link Team7 Tractogram SpaghettiBeans

   HAFT SimiaInuus TractographyValidationTeam TheUpsideDown BGKK

7. Results: ROC analysis TrackMcTrackface Accesschallenge X-link SimiaInuus HAFT

8. Results: ROC analysis TrackMcTrackface Accesschallenge X-link SimiaInuus HAFT TrackMcTrackface TwoPathsDverged

   Accesschallenge X-link Team7 Tractogram SpaghettiBeans HAFT SimiaInuus TractographyValidationTeam TheUpsideDown BGKK

9. Results: ROC analysis TrackMcTrackface Accesschallenge X-link SimiaInuus HAFT TrackMcTrackface Accesschallenge

   X-link SimiaInuus HAFT

10. Training Data Validation Data HCP DSI HCP DSI Results: Accuracy

    Comparison

11. Results: True Positives/False Negatives Localization

12. HCP DSI Results: True Positives/False Negatives Localization TPR TPR

13. • Better performance in training: tuning improves sensitivity Optimizing analysis

    for one injection site does not guarantee optimality for another Some methods showed consistently high performance in both datasets Conclusions

14. • Better performance in training: tuning improves sensitivity Optimizing analysis

    for one injection site does not guarantee optimality for another Some methods showed consistently high performance in both datasets • Higher sensitivity for DSI than multishell HCP acquisition scheme When analysis methods are optimized, the HCP scheme may achieve similar accuracy Conclusions

15. • Better performance in training: tuning improves sensitivity Optimizing analysis

    for one injection site does not guarantee optimality for another Some methods showed consistently high performance in both datasets • Higher sensitivity for DSI than multishell HCP When analysis methods are optimized, the HCP scheme may achieve similar accuracy • Probabilistic tractography + CSD model + use of constraining masks = increased sensitivity Best method for HCP acquisition scheme Conclusions

16. • Better performance in training: tuning improves sensitivity Optimizing analysis

    for one injection site does not guarantee optimality for another Some methods showed consistently high performance in both datasets • Higher sensitivity for DSI than multishell HCP When analysis methods are optimized, the HCP scheme may achieve similar accuracy • Probabilistic tractography + CSD model + use of constraining masks = increased sensitivity Best method for HCP acquisition scheme • False negative mainly found far from injections site and at splitting regions Conclusions

17. • What if all the participants applied some steps of

    the winner method? 2 winning strategies: post-processing filtering Use of constraining ROIs based on known anatomy Next Steps:

18. • What if all the participants applied some steps of

    the winner method? 2 winning strategies: post-processing filtering Use of constraining ROIs based on known anatomy Next Steps: IronTract ROUND II

19. • What if all the participants applied some steps of

    the winner method? 2 winning strategies: post-processing filtering Use of constraining ROIs based on known anatomy Next Steps: IronTract ROUND II • We are asking participants to: Run their tractography method on harmonized preprocessed data Post-process their data using the two winning post-processing strategies

20. • What if all the participants applied some steps of

    the winner method? 2 winning strategies: post-processing filtering Use of constraining ROIs based on known anatomy Next Steps: IronTract ROUND II • We are asking participants to: Run their tractography method on harmonized preprocessed data Post-process their data using the two winning strategies Monetary Prizes for winners! New Deadline: November 8th