An Examination of the Effect of Registration Errors on FDG-PET Evaluation of Chemotherapy Response in Sarcoma

Statistical quantitators that summarise standardized uptake value (SUV) distributions are widely used measures of metabolic activity in the analysis of static Positron Emission Tomography (PET) data. Amongst them, SUV mean and total lesion glycolysis (TLG) have been shown to yield a strong prognostic value for many diseases, which contributed to the rapid expansion of PET-based assessment of diagnosis, prognosis and treatment monitoring methodologies. Such measures, however, remain utilized in a point-wise fashion, that is, without a complementary assessment of their accuracy. Without such information, it is not clear that the assessment would be reliable. This raises important questions in particular for prognosis and therapeutic assessment. We consider here a statistical method that was recently proposed for the monitoring of neoadjuvant chemotherapy for sarcoma. This approach consists in pairing pre- and post-therapy scans in order to quantify the response to therapy, in terms of change in mean tracer uptake, along with an associated estimated accuracy. Pairing the uptake information requires co-registering the two sets of images, which most likely introduces a mis-registration error in the analytic framework. We propose to examine the effect of this mis-registration in a quantified analysis where we formulate the problem as a comparison with a classical (unpaired) therapeutic assessment. We demonstrate that mis-registration should not prevent paired-based methodologies. In particular, the derived measure of assessment accuracy remains more powerful even for large typical mis-registration scales. The viability of this method and its effect on prognostic utility are further considered via multivariate Cox survival analyses on a clinical dataset of 50 sarcoma studies with extensive follow-up information. Encouraging results suggest this approach could generalize to the analysis of other diseases and imaging modalities.