Maximum likelihood correction of shape bias arising from imaging protocol: Application to cardiac MRI

To establish a fair comparison between shape models derived from different imaging protocols, a mapping correcting local bias must be applied. In this paper, a multi-dimensional statistical model has been investigated to correct the systematic differences between Steady-State Free Precession (SSFP) and Gradient Recalled Echo (GRE) cardiac MRI protocols. This statistical model makes use of the Maximum Likelihood (ML) approach to estimate the local parameters of the respective GRE and SSFP distributions. Once those parameters are known, a local mapping can be applied. We applied this method to 46 normal volunteers who were imaged with both protocols. The SSFP model was estimated from the corresponding GRE model and validation was performed with leave-one-out experiments. The error was examined in both the local model parameters and the clinically important global mass and volume estimates. Results showed that the systematic bias around the apex and papillary muscles could be locally corrected and that the mapping also provided a global correction in cavity volume (average error of 0.4 ±12.4 ml) and myocardial mass (-1.2 ±11.1 g).