1. J OURNAL C LUB : Cardoso et al., University College London, UK “STEPS: Similarity and Truth Estimation for Propagated Segmentations and its application to hippocampal segmentation and brain parcelation.” Jul 15, 2013 Jason Su

2. Motivation Manual segmentation of structures on MR images is an important but often tedious task, requiring the expertise of a radiologist – Automation is highly desirable and can also help remove observer bias At 7T, we are working on studies that could benefit from this – Manually segmenting thalamic nuclei with the improved contrast of WMn-MPRAGE – Measuring atrophy in MS patients

3. Background STAPLE: Warfield et al. “Simultaneous Truth and Performance Level Estimation.” 2004. The algorithm considers a collection of segmentations and computes a probabilistic estimate of the true segmentation: “label fusion” Source of each segmentation in the collection may be human raters or automated True segmentation is formed by estimating an optimal combination of the segmentations, weighting each segmentation

4. Similarity Measures Intra-model: same contrast, same imaging parameters Mean Square Difference: 1/N*Σ(meanA-meanB) 2 Requires the images to have intensity values in the same range Normalized Cross Correlation (NCC): 1/N*cov(A,B)/(var A *var B ) Allows intensity values related by a linear transformation Inter-modal: Mutual Information: H(A,B) – H(A) – H(B) Indicates how much uncertainty about one set is reduced by the knowledge of the second set

5. Similarity Measures For binary or multi-class masks Dice coefficient 2[A == B]/[A] + [B] where [X] = the size of the region X

6. STEPS Improvements on STAPLE Use local intensity to select best labels to fuse Markov random field (MRF) to ensure spatial consistency can now handle probabilities Make unbiased towards larger structuresExtended these improvements to multi-label problem

7. Theory: Model For each voxel i: yi = image intensity t i = the true segmentation (binary or multi-class) d i = column vector of R candidate segmentations, i.e. our atlas found automatically or manually For each rater j: pj and qj represent how good she is at estimating 1s and 0s p j = P(d ij =1 | t i =1) q j = P(d ij = 0 | t i =0)

8. Theory: Model What we want: wi = P(ti=1 | di, p, q) The posterior probability Given all the information from our raters and how good they are, what is the likelihood that this voxel is actually in the structure Unfortunately p and q are coupled with wi We simultaneously modify our understanding of the truth and the performance of the rates compared to it This is solved iteratively with an Expectation-Maximization algorithm Alternatingly keep one fixed and solved for the other until convergence

9. Improvement: Local Ranking STAPLE did not have a full treatment of automatic segmentation with an atlas Accuracy of the registration was not taken into account Recently there have been developments using global or ROI metrics of NCC but these Introduced voxel-wise local NCC NCC computed within a Gaussian kernel around each voxel Algorithm is now modified to only keep top X registered raters per voxel Decouples sources of error, registration accuracy and reliability of rater

10. Improvement : MRF Regularization Spatial consistency to promote a contiguous segmentation is incorporated Uses Markov random field A voxel is independent of the rest of the image given its neighbors (6-connectivity) P(t i =k) is now also dependent on the voxel’s neighbors Adapted it to be computed with probabilities instead of the label, claim better accuracy and speed

11. Improvement: Unbiasing In STAPLE, the size of the segmentation/object heavily biases p and q If the object is small, we can guess 0s everywhere and still have a high rating for q Usually need to choose a high confidence threshold (~0.9999) to get the final segmentation Now only consider voxels in contention between raters Improves computation time and performance Choice of threshold is now simply 0.5

12. Validation Simulation Show the effect of local ranking using samples with different morphologies Compared against other algorithms with leave one-out cross validation in simulated MR data ADNI Database, measuring atrophy Evaluated improvement with each innovation Compared against other algorithms with leave one-out cross validation Other algorithms implemented in NiftySeg Where is STEPS?

13. Results: Optimizing Local Ranking STEPS does best for X=15 but the gain is marginal? Notably, other algorithms can become much worse with more templates

14. Results: Simulation Local ranking in STEPS beats global ranking in STAPLE Especially when the morphology of the target is very different from atlas raters This means can get away with less raters to cover more cases

15. Results: Phantom Simulation Measure the performance as R, the number of raters or size of the atlas, is reduced STEPS beats the others with even the smallest R – Performance characteristics can change with R, i.e. optimal X

16. Results: Phantom Simulation STEPS is significantly better than all of these competitors They all seem pretty good though, splitting hairs?

17. Results: Phantom Simulation

18. Results: ADNI

19. Results: MRF Smoothing in Multi-Steps

21. Discussion Each innovation in STEPS significantly improves the accuracy Final performance is significantly better than state of the art competitors Dice score (0.925±0.021) close to the inter-rater variability of the manual segmentors (0.93±0.03) on a different database Local ranking strategies implicitly encode local morphological variability rather than global Fewer anatomical templates are needed to deal with the population’s variability However, LNCC may have limitations in low contrast, consider other metrics