Saada A M Abujarada1, 2, James P B O'Connor1, 3, Chris J. Rose1, 2
1University of Manchester Biomedical Imaging Institute, Manchester, Greater Manchester, United Kingdom; 2University of Manchester Academic Health Science Centre, Manchester, Greater Manchester, United Kingdom; 3Department of Radiology, Christie Hospital NHS Trust, Manchester, Greater Manchester, United Kingdom
MRI-derived biomarkers, measured before treatment, may predict tumor response. Such research tends to use linear regression that may overlook differences in biomarker reliability (interscan repeatability). Where reliability is considered, it is often quantified separately to the regression task, rather than within the same conceptual framework. We describe a novel nonlinear model that explicitly models biomarker reliability and compare its predictions to those from a linear errors-in-variables model in a study of liver metastases treated with bevacizumab and FOLFOX6. While the median prediction error of the methods is equal, our method makes substantially fewer large errors.