302. Quantitative analysis and SVM-based classification of 18F-FDG PET data in mammary carcinoma

Abstract

Purpose Invasive ductal carcinoma is the most common type of breast cancer. The identification of the tumors with worst prognoses is important to achieve better clinical management of pathology. Aims of this study were the analysis of 18F-FDG PET radiometric features with Support Vector Machines (SVM) and correlation with prognostic clinical-pathological parameters. Methods The PET-CT scans of 51 women (age ⩾ 18 years) with invasive ductal mammary carcinoma selected for chirurgical resection were considered. Images were corrected voxel-by-voxel for Partial Volume Effect using an iterative post-reconstruction method based on measured system Point Spread Function. Mean Standardized Uptake Value (SUVmean) was extracted for segmented Regions Of Interest (ROIs) representing tumor lesions (L) and background mammary healthy tissue (BK). Pearson correlations of tumor proliferative index with SUVmean(L) and SUVmean(L) normalized to SUVmean(BK) (SUVmean(L)-norm) were computed. The relevance of SUVmean(L)-norm and SUVmean(BK) to discriminate tumor luminal subtypes (A, B, B2, HER2, TN) [1] was investigated implementing linear-kernel SVM classification [2] . Performance was evaluated computing the Area (AUC) Under the Receiver Operating Characteristic (ROC) curve [3] , estimated in leave-one-out cross-validation to ensure an unbiased estimate of the classifier performance. Results Significant positive correlations of SUVmean(L) (p = 0.007) and SUVmean(L)-norm (p = 0.003) with proliferative index were found. Considering the features SUVmean(L)-norm and SUVmean(BK), multiclass classification was not able to discriminate between all luminal categories with performances significantly different from the chance level. By contrast, considering two-class classification, AUC was enhanced to 81% when considering luminal TN vs others and 89% for luminal A vs others. Conclusions Quantitative features extracted from PET images were correlated with prognostic clinical and histopathological factors in mammary carcinoma. Moreover, the SVM-based classification demonstrated a predictive value of PET images.