BootSMT: Bootstrap Split/Merge Fiber Tractography for DT-MRI

LL-PHS-WE5A

BootSMT: Bootstrap Split/Merge Fiber Tractography for DT-MRI

Scientific Informal (Poster) Presentations

Presented on November 30, 2011
Presented as part of LL-PHS-WE: Physics

Ahmet Burak Yoldemir, Abstract Co-Author: Nothing to Disclose

Burak Acar PhD, Presenter: Nothing to Disclose

To develop a technique to improve and assess the reliability of fiber tract sets via estimating pairwise connectivity probabilities of short (more reliable) tracts generated from the given set by splitting.

Fiber tracts computed by 4th order Runge-Kutta (RK4) method are split into short tracts {S} at low fractional anisotropy points. Ck={(Sm , Sn, Jmn)} defines a cluster of S’s, sequentially connected by bridges J. Jmn’s fitness is derived from the diffusion tensors at the connected end-points and the collinearity of Sm and Sn. Bootstrap is used to sample the space of Ck’s. Taking each Sm as the seed, initial cluster Cm0 is formed by recursively selecting the most fit bridge at the free end-point of the most recent S connected. Bootstrap sample Cmk’s are formed by probabilistically replacing a bridge of Cm0. Unweighted bootstrap sampling is applied to these cluster sets to estimate the distribution of Ck’s. Pairwise connectivity probabilities between the seed tract Sm and others, Pn|m=P(Sn|Sm), are estimated in Bayesian sense from the estimated distribution of Ck’s. Short tract clusters are interactively formed for given seed tracts and varying thresholds as Tm(t)={ Sn | Pn|m ≥ t}. Quantitative analysis is performed on phantom DT-MRI dataset [http://cubric.psych.cf.ac.uk/commondti/]. For t ∈ [0,1], sensitivity (SN) of Tm(t) is defined as the percentage of the length of the true tract Γm, covered by ({Sn}∈ Tm(t) ∧ {Sn}∈ Γm) while the false positive (FP) rate is defined as the total length of ({Sn}∈ Tm(t) ∧ {Sn}∉ Γm). RK4 tractography is evaluated likewise.

8 seed locations are determined uniformly for each geometry and used for all (SNR=5/15/30). SN’s @ FP=0 for (RK4:bootSMT) are: Ortho-crossing: (0.35±0.12:0.61±0.23/ 0.37±0.10:0.87±0.14/0.58±0.23:0.95±0.10) and corresponding t’s are (0.44±0.28/0.10±0.17/ 0.01±0.01). Kissing: (0.34±0.25:0.72±0.25/ 0.47±0.23:0.97±0.10/0.61±0.11:1.00±0.01) and corresponding t’s are (0.26±0.33/0.07±0.16/ 0.08±0.15).

We have demonstrated quantitatively that bootSMT improves the sensitivity of RK4 on average by 92.2% without increasing FP rate. Decreasing t with increasing SNR indicates that bootSMT is sensitive to data quality as well as providing an interactive interface to investigate the reliability of tractography results.