Figure 3
Accuracy and precision of model fitting. (A) The mean (data points) and variance (error bars) of the difference between the prediction for DNN or conventional non-linear least squares optimization and the ground truth values are plotted against the known ground truth from numerical simulations. The line at zero difference is also plotted as straight black line to aid appreciating the accuracy of the prediction from both methods (higher the accuracy, closer the mean difference to zero). The variance of the difference (error bars) is a good indicator of the precision of the estimation: smaller the variance, higher the precision. To make the results visually clear, data points for the non-linear least squares were purposely moved slighted to the right. (B) The probability density distribution of the estimates of the seven rVERDICT model parameters (S0 was fixed to 1) are plotted for seven ground truth values representative of PCa in TZ and PZ and 4096 different random realizations of the other parameters, for both DNN and conventional non-linear least squares optimisation. The wider the distribution, the less robust the estimation and the lower the precision due to degeneracy and/or spurious minima. We quantified the width of the distributions through their standard deviations (S.D.) reported in each plot with corresponding matching colours.
