Robustness and Benefits of Longitudinal Image Co-Registration for Prediction of Triple Negative Breast Cancer Response to Neoadjuvant Chemoimmunotherapy
Abstract
Purpose
To establish a co-registration pipeline that achieves robust alignment of multiparametric breast MRIs acquired at multiple visits, improving accuracy in assessment of triple negative breast cancer (TNBC) response to neoadjuvant chemoimmunotherapy (NACI).
Methods
This study included 29 early-stage TNBC patients (14 pathologic complete response [pCR], 15 non-pCR) from the ARTEMIS trial (NCT02276443) who received standard-of-care NACI and underwent longitudinal multiparametric MRI acquisition (baseline, after 2 cycles of Paclitaxel+Pembro, after 4 cycles of Paclitaxel+Pembro). Tumor contours were segmented by two radiologists. Imaging data were processed through our longitudinal analysis pipeline (Jarrett et al. Nature Protocols 2021). Intra-visit rigid co-registration was first performed to account for patient motion and misalignment between DCE- and DWI-MRI. Afterwards, inter-visit co-registration used b-spline nonrigid registration with a rigidity penalty on the tumor region. The second visit was used as a reference and the whole breast was aligned without biasing tumor shape and size. We calculated the tumor volume and longest-axis dimension before and after inter-visit co-registration to 1) confirm pipeline robustness in preserving 3D tumor morphology, evaluated by concordance correlation coefficient (CCC) between the pre and post co-registration “delineated tumor volumes”, and 2) investigate whether longitudinal registration improved the prognostic value of the “longest-axis” and “clinical volume” (calculated from the axis lengths of tumor in 3 dimensions), quantified by ROC AUC for differentiating pCR status using pre versus post co-registration measures.
Results
Tumor volumes were well preserved, with CCC of 98.7% after co-registration. ROC analysis using the longest axis yielded AUC of 0.781 pre-coregistration and 0.800 post-coregistration. Clinical tumor volume ROC analysis yielded AUC of 0.748 pre-coregistration and 0.752 post-coregistration. Both indicated improvement in prediction power.
Conclusion
Our framework effectively retains the tumor's structural characteristics while correcting orientation bias and improving tumor progression assessment. Ongoing efforts include employing a larger cohort and performing habitat analyses.