In clinical studies in which medical imaging is used as an aid in the detection of disease, or of disease progression, interesting questions arise regarding the noise in the imaging endpoints (reflected, say, in reader discordance) and its effect on the classification of disease status. In this presentation, we describe some of the issues and in some cases, propose possible ways to move toward their resolution.

A particular problem we will touch upon is as follows: In a cancer clinical trial, a biomarker - a specified increase in tumor size – is substituted for a hard endpoint – death. The measurements of change in tumor size are “noisy”, as they are read from images. This creates a serious problem, since it’s then possible that the calculated survival curves could differ significantly from the true survival curves and therefore lead to biased estimation of treatment effect.