MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Feasibility of Whole Brain Voxel-Wise Statistical Analysis of Longitudinal Diffusion Tensor Imaging (DTI) of Single Subject: Quantitative Comparisons of Three Non-Parametric Methods

T Zhu1*, X Qiu2 , J Lian3 , S Das4 , (1) university of North Carolina at Chapel Hill, Chapel Hill, NC, (2) University of Rochester, Rochester, NY, (3) university North Carolina, Chapel Hill, NC, (4) University of North Carolina, Chapel Hill, NC

Presentations

(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Karl Dean Ballroom C

Purpose: Conventional statistical comparisons are designed to detect common changes at same spatial locations within patient group/population, therefore unlikely to detect subject-specific radiation-induced brain injuries (RIBIs) from brain RT. This study evaluated three non-parametric methods, permutation, wild-bootstrap and spatial regression, which enable statistical inference of longitudinal DTI for a single subject to detect RIBIs.

Methods: Statistical power was evaluated through Monte Carlo (MC) simulations of longitudinal DTI and longitudinal DTIs of one patient with active lesions. For MC, two synthetic brain DTIs (pre- vs. post-RT) were created. Within a 5x5x3 region at corpus callosum, a synthetic lesion was simulated by differing the largest eigen values of two DTIs to a certain amount. Four difference levels (=10%/20%/40%/80%) were evaluated, simulating different severity levels. 100 MC resamples were created for each method. Receiver-Operating-Characteristic (ROC) curves were then generated from true and false positive ratio at six different p values (=0.01/0.5/0.1/0.15/0.2/0.3). For longitudinal DTI of a patient with progressive active injuries, detected lesions by three methods were compared against manual lesion contour by an experienced neuroradiologist.

Results: ROC analyses from simulations show that wild-bootstrap provides the best sensitivity (=0.31 even with a difference level of 10%) and permutation provides the best specificity (=0.99 with a difference level of 80%). Spatial regression provides good combination of sufficient sensitivity and high specificity. For the patient, all methods detected significantly increased mean diffusivity (MD) of active lesions. Spatial regression and permutation achieved good balance between sensitivity and specificity.

Conclusion: Statistical inference can be achieved for longitudinal DTI analysis of the same individual, even when only one DTI were acquired at each time point. With good control of false positive detections, both spatial regression and permutation methods can potentially provide practical approaches to assess individual's radiation-induced brain injury by establishing a within-subject baseline and constructing data-drive distributions.

Keywords

MRI, Statistical Analysis, Radiation Therapy

Taxonomy

Not Applicable / None Entered.

Contact Email