Over the last year we have been working on a new generalized linear-combination modeling in Osprey. Generalized means that every modeling decision can be controlled by the user through a single interface json file. The implementation also allows for modeling of arbitrary 2D experiments.

In the first study, we used benchmarked the new algorithm by designing a simple synthetic MRS dataset representing a multi-transient conventional MRS experiment of a single metabolite (Scyllo-Inositol and GABA). The data was modeled using traditional methods (averaging and 1D-LCM) and multi-transient 2D-LCM of all averages at once. Bias in concentration estimates agreed well between both methods. Interestingly we found a slight reduction in the variance of the CRLBs for 2D-LCM compared to traditional 1D-LCM. In case of correlated noise, 2D-LCM CRLB showed a significant improvement and higher agreement with the ground-truth CRLBs.

All results and code are available on OSF.

Links