Abstract
438
Objectives In this study, we evaluated the clinical value of intra-tumor heterogeneity as a novel biomarker on the assessment of early therapeutic response using 18F-FDG PET. Support Vector Machine (SVM), a novel machine learning tool, was applied to improve the accuracy of prognosis.
Methods Twenty-five patients with oral cavity squamous carcinoma underwent 18F-FDG PET/CT scans before the onset of concurrent chemo-radiotherapy (CCRT) and after the completion of first cycle (7 days gap between two scans). The 5-year survival was used as gold standard to categorize the patients into 2 groups: responders (n=16) and non-responders (n=9). We applied the global Moran I(d) analysis to characterize the intra-tumor heterogeneity on PET images @1h post-injection. Other than conventional texture analysis that widely used in heterogeneity prediction, I(d) statistic is a measure of spatial correlation among 3D neighboring voxels in the target volume of interest normalized by the autocorrelation. We compared the performance of image features including the heterogeneity and SUVmean & max in differentiating the responsive patient group from the non-responsive group. For each image feature, the quantitations derived from the PET images before and after the initiation of the treatment were tested as individual biomarkers. The student t-test was performed to compare the responders vs. non-responders. Meanwhile, receiver operating characteristic (ROC) was used to characterize the prognosis accuracy of the prediction using each image feature. Then we trained and applied a SVM based machine learning tool to fuse the heterogeneity I(d)s derived from pre and post-treatment scans for prospective prediction. The trained structure was used to test the SVM performance in classifying responsive/nonresponsive patient groups using cross-one-validation.
Results Both individual intra-tumor heterogeneities I(d) derived from PET images before and after the initiation of treatment can significantly differentiate the non-responders from responders ( Pre-treatment: non-responders 0.69 ± 0.11 vs responders 0.39±0.2, P=0.0024; Post-treatment: non-responders 0.54±0.13 vs responders 0.3 ± 0.2, P=0.012). Neither SUVmean nor SUVmax demonstrate the similar capability. With regard to the percentage changes between pre- and post-treatment images, none of them (heterogeneity, SUVmean and SUVmax) show significant difference between the two patient groups judging by student t-test. ROC analysis confirms that heterogeneity yielded effective prognosis of the treatment (AUC 89% pre, 82% post) comparing with SUVmax (AUC 62% pre, 58% post). Interestingly, the best non-responder/responder classification was achieved when both intra-tumor heterogeneities I(d) derived from pre and post-treatment PET images were fed in Support Vector Machine as dual-input (accurate rate 90.48%), which is superior than using each of them as single input ( 86.7% for I(d) pre, 71.4% for I(d) post respectively). Neither the individual SUVs nor their combinations effectively differentiated the responders from non-responders using SVM classification (accurate rate of SUVmax pre 53%, post 57%, combined 63%).
Conclusions Intra-tumor heterogeneity showed superior performance than other conventional image features when single feature was applied to early treatment evaluation. Machine learning based nonlinear classifier such as SVM can achieve very encouraging prognosis when including both intra-tumor heterogeneities derived from PET images before and a week after initiation of the treatment.