MRI-based Tumor Habitat Analysis for Treatment Evaluation of Radiotherapy on Esophageal Cancer

Abstract

Introduction: We aim to evaluate the performance of pre-treatment MRI-based habitat imaging to segment tumor micro-environment and its potential to identify patients with esophageal cancer who can achieve pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT). Material and methods: A total of 18 patients with locally advanced esophageal cancer (LAEC) were recruited into this retrospective study. All patients underwent MRI before nCRT and surgery using a 3.0 T scanner (Ingenia 3.0 CX, Philips Healthcare). A series of MR sequences including T2-weighted (T2), diffusion-weighted imaging (DWI), and Contrast Enhance-T1 weighted (CE-T1) were performed. A clustering algorithm using a two-stage hierarchical approach groups MRI voxels into separate clusters based on their similarity. The t-test and receiver operating characteristic (ROC) analysis were used to evaluate the predictive effect of pCR on habitat imaging results. Cross-validation of 18 folds is used to test the accuracy of predictions. Results: A total of 9 habitats were identified based on structural and physiologic features. The predictive performance of habitat imaging based on these habitat volume fractions (VFs) was evaluated. Students’ t-tests identified 2 habitats as good classifiers for pCR and non-pCR patients. ROC analysis shows that the best classifier had the highest AUC (0.82) with an average prediction accuracy of 77.78%. Conclusion: We demonstrate that MRI-based tumor habitat imaging has great potential for predicting treatment response in LAEC. Spatialized habitat imaging results can also be used to identify tumor non-responsive sub-regions for the design of focused boost treatment to potentially improve nCRT efficacy.