Applying biophysical models to diffusion MRI data has the potential to provide more specific information about tissue microstructure, beyond indices such as the apparent diffusion coefficient (ADC)^1,2. In particular, where different microstructural changes result in similar ADC changes, the use of biophysical models may allow these situations (e.g. change in cell size vs. change in volume fraction) to be distinguished. However, fitting such models to noisy data can yield imprecise and inaccurate parameter estimates with artificial correlations between parameters³, potentially hampering the utility of such approaches. Here, optimum experimental design concepts^4,5 are used to investigate the effects of acquisition parameters on estimates from a simple model of tumour tissue undergoing distinct microstructural changes. Optimum acquisitions are found, and used in model-fitting simulations, comparing results with those obtained using non-optimum acquisitions.Model: The normalised PGSE signal, S/S₀, was modelled analytically by combining restricted diffusion inside a sphere with hindered extracellular diffusion, with four model parameters: cell radius, R, intracellular volume fraction, f­_i, and intra- and extra-cellular diffusivities, D_i, and D_e; S₀=1. Starting from a ‘baseline’ with R=10 µm, f_i=0.6, D_i=D_e=1.5 µm²/ms, representing a plausible model of tumour tissue, two possible microstructural changes were considered: (1) a decrease in f­_i to 0.24 (a cell density decrease, mimicking complete cell death), (2) a decrease in R to 7.37 µm with an associated decrease in f_­i to 0.24 (cell density remains constant, mimicking apoptotic cell shrinkage, with a 60% cell volume decrease⁶). Changes (1) and (2) both give an ADC increase for a typical multi-b-value clinical protocol; see Figure 1, top. Optimum design: Optimum PGSE parameters (gradient strength, G, separation, ∆, and duration, δ) are those that maximise or minimise some summary statistic of the signal model’s information matrix, M^4,5. Sensitivities for M (∂S/∂R, ∂S/∂f_i, ∂S/∂D_i, and ∂S/∂D_e) were calculated numerically for 57016 combinations of {G, ∆, δ} satisfying typical clinical constraints: G_max=60 mT/m, TE_max=100 ms, b_min=150 s/mm² (avoiding perfusion effects). Sensitivities were used in a Fedorov exchange algorithm⁷ to find four optimum {G(mT/m), ∆(ms), δ(ms)} combinations for estimating R, f_i, D_i, and D_e for each microstructure. D-optimum designs⁴ were calculated, corresponding to maximising the determinant of M. Simulations: Synthetic signals were generated for all three microstructures, for three acquisition strategies: (a) D-optimum for each specific microstructure, (b) D-optimum for the ‘baseline’ microstructure, and (c) non-optimum, taken as {30,20,15}, {30,72,15}, {60,20,15}, {60,72,15}; i.e. fixed δ, variable G and ∆. Gaussian noise (SD=S₀/SNR, 5000 instances, SNR=10⁶, 50) was added, and the model was fit using least-squares.Figure 1 (bottom) lists D-optimum designs for each microstructure. Figure 2 plots the correlations between the model parameters for the ‘baseline’ microstructure at SNR=10⁶ for non-optimum and D-optimum designs. The D-optimum acquisition tends to reduce the correlation between parameters (most evident on the top row) and improves precision. Behaviour at a more realistic, but still reasonably high, SNR of 50 (Figure 3), shows that while similar correlation patterns are observed for the two designs, there is benefit in using the D-optimum acquisition: the second peak of the R distribution is suppressed, D_i is estimated with better precision, and D_e and f_i­ estimates are more accurate. Figure 4 shows R and f_i histograms for the three acquisition strategies, for each microstructure. The main benefit of D-optimum designs appears to be in R estimates: distributions for the two changes are similar and broad when using non-optimum designs, while D-optimum designs considerably improve the precision, offering a greater chance of detecting the change in cell size. This is true for the best-case scenario where each microstructure is assessed with its own specific D-optimum design (Figure 4, middle column), as well as the more realistic scenario where a single optimised protocol is used for all microstructures (Figure 4, right column). Note that the change in R considered here is relatively large⁶ and sensitivity to smaller changes will be lower, especially as SNR decreases. Precision of f_i estimates is generally improved with D-optimum designs, while D_i estimates are generally poor for each strategy and D_e distributions are similar (D_i and D_e results not shown).D-optimum designs of diffusion MRI acquisitions may prove beneficial for distinguishing between different changes tumour tissue may undergo. Note that some knowledge of likely underlying microstructure is required to enable optimisation, and further work will assess the extent to which a single optimised protocol can be applied to a wider range of physiologically relevant microstructures. Joint consideration of optimised acquisition parameters and the magnitude of specific biological changes is likely to be important when evaluating the utility of microstructural models.