부산대학교 도서관

In this project, we assessed the clinical value of tumor heterogeneity measured with 18 F-FLT as a biomarker for lung cancer diagnosis and staging, then compared its performance to traditional image features using final pathology as gold standard. We also proposed to apply support vector machine (SVM) to train a vector of image features including heterogeneity extracted from PET image and CT texture features to improve the diagnosis and staging for lung cancer. Thirty-two subjects with lung nodules (19 M, 13 F, age 70 ± 9 y) who underwent 18 F-FLT PET/CT scans were included in our study. We applied the global Moran I(d) analysis to characterize the intra-tumor heterogeneity on PET images 1h post-injection. Other than texture analysis that widely used in heterogeneity prediction, I(d) statistic is a measure of spatial autocorrelation characterized by the correlation among 3D neighboring voxels. Other image features including SUV and CT texture were extracted from PET/CT images. Then we trained and applied a SVM based statistical machine learning tool to fuse the features and test the SVM performance in classifying patient groups: benign / early malignant and early / advanced malignant. Heterogeneity derived from 18 F-FLT images significantly differentiated benign (0.24 ± 0.09, N=9) from early stage malignancy (0.40 ± 0.09, N=10; P = 0.002), as well as early stage from advanced stage malignancy (0.50 ± 0.07, N=13, P = 0.005). Other image features, SUVmean and CT texture, didn't demonstrated similar capability. Intra-tumor heterogeneity showed superior performance than other traditional image features when single feature was applied to staging. Furthermore, the SVM classification showed that best performance of staging was achieved when all image features are combined in the SVM training. In conclusion, we obtained a novel measurement of intra-tumor heterogeneity which has promising performance for diagnosis and staging of lung cancer. We demonstrated the feasibility of performing SVM based cancer staging using multiple image features in PET/CT. SVM analysis and classification with combination of effective features has great potential to augment diagnostic accuracy and improve tumor staging in oncological practice.