Purpose: Quantification of local metabolic tumor volume changes (âˆ†MTV) after chemo-radiotherapy (CRT) would allow accurate tumor response evaluation. The purpose of this study is to measure local âˆ†MTV using blended PET/CT registration and to provide an early prediction of pathologic tumor response.
Methods: 61 patients with locally advanced esophageal cancer underwent baseline, post-induction chemotherapy (follow-up) and post-CRT PET/CT. A grayscale blended PET/CT image was generated and follow-up blended PET/CT image was registered to the baseline blended PET/CT image with tumor motion correction. Jacobian map (J) was computed from the transformation which measured local MTV shrinkage (J < 1) or expansion (J > 1). Registration accuracy was evaluated by comparing the net âˆ†MTV calculated by Jacobian integral [(mean J â€“ 1) x baseline MTV] vs. semi-automatic segmentation. Radiomic features were then extracted from the Jacobian maps and PET/CT images and distinctive features were identified using a hierarchical clustering method. A logistic regression model was constructed using two distinctive features pre-identified from PET/CT images (PET Kurtosis) and Jacobian map (Mean of Cluster Shade) to predict pathologic complete response (pCR). The model was evaluated using 10x5-fold cross-validation.
Results: Qualitatively, Jacobian map by blended PET/CT registration showed smoother local âˆ†MTV than PET-PET and CT-CT registrations. Quantitatively, âˆ†MTV calculated by Jacobian integral of blended PET/CT registration (-42.0%) was closer to âˆ†MTV measured by the semi-automatic segmentation (-50.0%) than by PET-PET (-29.4%) and CT-CT (-8.0%) registrations. Kurtosis in follow-up PET and Jacobian Mean of Cluster Shade represented uniformity of FDG uptake and asymmetry of âˆ†MTV, respectively and the logistic model built with these two features achieved a high accuracy (AUC=0.82) in predicting pCR.
Conclusion: The novel Blended PET/CT registration led to more accurate quantification of âˆ†MTV than PET-PET and CT-CT registrations. The model with combination of PET/CT and Jacobian features showed high accuracy in predicting pCR.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part through the NIH/NCI Grant R01CA172638 and the NIH/NCI Cancer Center Support Grant P30 CA008748.