Quantitative Assessment of Volume Change in Tumors Using Image Registration.

Abstract

Assuming tumor volume change will be shown to be a biomarker for therapeutic response, accurate early quantification of tumor volume change could lead to interactively adapting an individual patient’s therapy such as drug or dose modification to achieve optimal response as well as shorter phase III clinical trials. Standard techniques for response evaluation like RECIST 1.1 or WHO measurement methods have limited effectiveness in accurately assessing small, early changes. Most current approaches estimate a tumor's change indirectly by independently segmenting it in interval exams and then subtracting the segmented volumes to obtain a change estimate. Ensuring the consistency of these independent segmentations across interval exams can be a significant challenge. This thesis develops a low noise, low bias direct algorithm to measure volume change using 3D image registration. Tumor pairs are spatially registered across intervals and volumetric change is calculated by summing local scale changes obtained from the Jacobian map of the deformation. Such an approach can also potentially show regions of differential growth and contraction across the lesion. The registration-based algorithm is evaluated using synthetic and in-vivo interval scans where true tumor volume change is unequivocally known. The 95% confidence error interval of measured volume change was (-8.93%, 10.49%) and (-7.69%, 8.83%) using mutual information and normalized cross correlation, respectively, as similarity measures for registration. To the best of this author’s knowledge, these are the tightest bounds reported thus far for zero-change in vivo studies. No statistically significant evidence of functional bias was found for the registration-based volume change measurement algorithm. Statistical models are developed to show that using the registration-based algorithm the error in measuring volume change increases with increase in tumor volume and decreases with the increase in tumor's normalized mutual information, even when that is not the similarity measure being optimized. The developed registration-based algorithm is also compared with other approaches to demonstrate that it has the potential to outperform indirect segmentation-based change measurement methods. The potential of an accurate registration-based change measurement algorithm in tracking progression of chronic obstructive pulmonary disease is also suggested through an initial study on a normal and a diseased patient.