Purpose: To compare the theoretical information content of four popular models of diffusion-weighted signal attenuation.
Method: Four whole prostates were imaged fresh unfixed and fixed at 9.4T. Biexponential, kurtosis, stretched exponential, and monoexponential models were ranked using Akaike's Information Criterion (AIC) with validation by a leave-one-out test of model prediction error.
Results: For unfixed tissue measurements (b-value range: 17-2104 s/mm(2)) the biexponential and kurtosis models had similar information content to each other and this was distinctly higher than for the stretched and monoexponential models. In fixed-tissue measurements (b-value range: 17-8252 s/mm(2)), the biexponential model had much higher information content than the three other models.
Conclusion: AIC-based model ranking is consistent with an independent prediction accuracy test. Biexponential and kurtosis models consistently perform better than stretched and monoexponential models. The biexponential model has increasing superiority over all three other models as maximum b-value increases above ∼2000 s/mm(2).
Keywords: Akaike Information Criterion; diffusion; modelling; prostate.
Copyright © 2013 Wiley Periodicals, Inc.