I’m aware that you’re working on implementing CRLBs. I’m wondering if there is a way of assessing model fit with the current outputs? Is the residual comparable across participants/sequences (i.e. if the number for the residual is a lower I can assume the model is a better fit?).
Also, is there a reason you switched from using the NAA peak for SNR to the Creatine peak?
The currently available metric is the relative residual which is stored in the QA tsv termed relRes[Subspec]. The lower the number the better the agreement between data and model. For the current model this is a relatively good indicator for fit quality. However, it doesn’t tell you anything about how reliable a specific metabolite is estimated.
The metric should be comparable between subjects and the same sequences. However, it will be different for the off and the diff spectrum for example.
We switched from NAA to creatine because we are part of a multi-center infant study where NAA is lower compared to the adult brain. Creatine on the other hand is comparably stable allowing us to better compare data quality between adults and infant data.
The SNR is calculated with FID-A native functions (osprey/op_getSNR.m at develop · schorschinho/osprey · GitHub). It uses the peak maximum in the defined range (in our case creatine) and the SD of the detrended noise (-2 to 0 ppm). In addition any DC offset is removed from the peak.
We are using ONE SD not 2 as it is recommended in the terminology consensus paper.