TY - JOUR T1 - Radiomics features of <sup>18</sup>F-FDG PET/CT predicting breast cancer molecular subtype: a preliminary study<strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 555 LP - 555 VL - 61 IS - supplement 1 AU - Xiaojun Xu AU - Yongxue Zhang AU - Jinxia Guo AU - Xiaoli Lan Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/555.abstract N2 - 555Purpose: The molecular subtyping of breast cancer is closely related to the therapeutic strategy and patient prognosis. This study aimed to investigate the feasibility of predicting molecular subtypes of breast cancer by imaging radiomics features extracted from 18F-FDG PET/CT images before treatment. Methods: We retrospectively analyzed breast cancer patients who underwent 18F-FDG PET/CT examinations from January 2012 to December 2018. Inclusion criteria were as follows: (1) pathological diagnosis was breast cancer and molecular subtype was given; (2) no treatment before PET/CT imaging; (3) complete clinical data. Exclusion criteria included: (1) no pathological molecular subtype; (2) chemotherapy, local puncture biopsy, surgery or other treatments before PET/CT imaging; (3) history of other malignancies; (4) blood glucose level was more than 11.1 mmol / L. Clinical parameters, including age, tumor size, initial T, N, and M category, pretreatment CEA, CA125, CA153 were also collected. Regions of interest (ROI) were drawn manually on PET and CT images. We used IF Foundry (Intelligence Foundry 2.1, GE Healthcare) to extract functional imaging parameters (SUVmax and SUVmean, metabolically active tumor volume [MTV] and total lesion glycolysis [TLG], CT HUmean and CT Volume) and radiomics features (histogram, shape, textural and contour features, Intra-perinodular features [Ipris], co-occurrence of Local Anisotropic Gradient Orientations features, and filter-based features including Wavelets, Gabor, local binary pattern [LBP] and Wavelets + LBP). Modeling was conducted by SoftMax multi-class logistic regression, calculated by the three-fold cross-validation, and the results were evaluated with accuracy (ACC). Multi-class logistic regression modeling applied two strategies, One-Versus-One (OVO) and One-Versus-Rest (OVR), and selected different random seeds. The three-fold cross-validation was repeated 500 times to obtain the ACC distribution histogram and average. Multi-classification studies were performed on six combinations of CT, PET, PET/CT, PET/CT + functional imaging features, PET/CT + biomarker features, PET/CT + functional imaging features + biomarker features. Results: Eighty female patients (age range, 25-81y; average, 53.0 y) were included in this study, including 65 cases of invasive ductal carcinoma, 6 cases of invasive lobular carcinoma, 3 of ductal carcinoma in situ, 3 cases of mixed invasive carcinoma, 2 cases of mucinous carcinoma, and 1 case of medullary carcinoma. Patients were divided into 4 subgroups (Luminal A, 19 cases; Luminal B,30 cases; Her-2 positive,16 cases; and triple negative, 15 cases) according to the molecular subtyping. After screening, the radiomics features included 21 CT and 19 PET dimensions. The 21 dimensions of CT features were original 1, textural 1, CoLIAGe2D 5, wavelets + LBP 4, Gabor 6, PLBP 1, WILBP 3. The 19 dimensions of PET features were CoLIAGe2D 1, Wavelets + LBP 4, Gabor 12, PLBP 1, WILBP 1. The accuracy of PET and CT radiomics features for single-mode and multi-mode modeling classification were shown in the Figure. The results of PET/CT radiomics features (OVR and OVO) were better than those of PET or CT. Combined biomarkers and functional imaging features with PET/CT features could improve the diagnostic efficacy, but no statistical difference exists. Compared with OVO, better results are achieved using the OVR strategy for modeling. Conclusions: PET/CT radiomics features could provide predictive value for breast cancer molecular subtyping. More research with large sample size needs to be performed to confirm this. This work was supported by National Natural of Science Foundation of China (No. 81630049) ER -