Stefan Klein
1 like430 views
This document discusses an approach to early diagnosis of dementia based on comparing brain images. It involves calculating the distances between brain images rather than extracting predefined features. The distances represent how dissimilar the brain anatomies are. The approach was tested on brain MRI data from 490 subjects. Using only image information, it could correctly classify 80% of whether subjects who were healthy at baseline developed dementia 10 years later, outperforming classification based solely on age. This dissimilarity-based method shows promise as an approach for early dementia prediction based on brain image information.
Stefan Klein
- 1. Early diagnosis of dementia based on intersubject whole-brain dissimilarities Stefan Klein Biomedical Imaging Group Rotterdam (BIGR)
- 2. Research question: 1 Will the owner of this brain become demented?
- 3. Common pattern recognition approach: Assemble a training set 2
- 4. Common pattern recognition approach: 1 Assemble a training set
- 5. Common pattern recognition approach: Assemble a training set Calculate features Hippocampus volume CSF volume Number of WML Cortex thickness ... feature 1 feature 2 x x x x x 3
- 6. Common pattern recognition approach: Assemble a training set Calculate features Hippocampus volume CSF volume Number of WML Cortex thickness ... 3. Train a classifier feature 1 feature 2 x x x x x 4
- 7. Dissimilarity based approach: Assemble a training set 5
- 8. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value 5
- 9. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value x x x x x x x x x x
- 10. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value x x x x x x x x x x 0 0 0 0 0 0 0 0 0 0
- 11. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value x x x x x x x x x x 0 0 0 0 0 0 0 ... 0.2 0.1 0 0.3 0.5 0 0.2 0
- 12. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value x x x x x x x x x x 0 0 0 0 0 0 0 ... 0.2 0.1 0 0.3 0.5 0 0.2 0
- 13. Dissimilarity based approach: Assemble a training set Calculate distances between subjects! similar anatomy small distance value Train classifier on distance matrix x x x x x x x x x x 0 0 0 0 0 0 0 ... 0.2 0.1 0 0.3 0.5 0 0.2 0
- 14. Distance between images How ? 6
- 15. Nonrigid image registration fixed image moving image 7
- 16. Nonrigid image registration fixed image moving image 7
- 17. Nonrigid image registration fixed image moving image use “complexity of deformation field” as distance measure! 7
- 18. Nonrigid image registration fixed image moving image use “complexity of deformation field” as distance measure! Distance = std.dev. ( det ( jacobian ( deformation field ) ) ) 7
- 19. Experiment MR brain data of Rotterdam Scan Study (1.5T MR HASTE) 490 subjects: all healthy when scanned (’95/’96) 10 years later: 45 subjects developed clinical symptoms of dementia Randomly select 45 subjects of the remaining cases: healthy controls. 8
- 20. Experiment MR brain data of Rotterdam Scan Study (1.5T MR HASTE) 490 subjects: all healthy when scanned (’95/’96) 10 years later: 45 subjects developed clinical symptoms of dementia Randomly select 45 subjects of the remaining cases: healthy controls. = training & test set (leave-one-out cross-validation) 8
- 21. 9
- 22. 80% correctly classified 9
- 23. However:
- 24. However: Based on age alone: 69% correctly classified 10
- 25. However: Based on age alone: 69% correctly classified USE AGE-MATCHED CONTROLS!!! 10
- 26. Conclusions Dissimilarity based classification is very promising. Dementia could be predicted with a high accuracy, based on only image information. Use age-matched controls!! 11
- 27. Research question: 11 Will the owner of this brain become demented?
- 28. Research question: Will the owner of this brain become demented? Answer: YES! 11
- 29. Acknowledgements Marco Loog Bob Duin Alexander Hammers Fedde van der Lijn Tom van den Heijer Marleen de Bruijne Aad van der Lugt Monique Breteler Wiro Niessen 12
- 30. Developed together with Marius Staring: Rigid and nonrigid registration Various cost functions (SSD, mutual information, etc) Fast because of stochastic optimisation Free: http://elastix.isi.uu.nl Based on Insight ToolKit (ITK): http://www.itk.org
- 31. Application: brain segmentation Atlas matching (Atlas = MR T1, multimodal registration required)