Breast cancer subtype has to be accounted for to predict response to neoadjuvant chemotherapy based on radiomic features in 18F-FDG PET

1236 Objectives: Breast cancer (BC) is a heterogeneous disease and various BC subtypes account for part of this heterogeneity. We investigated whether radiomic features (RF) calculated from FDG PET also reflected BC heterogeneity and determined whether RF combined with other important BC prognostic factors could predict pathological complete response (pCR) in BC patients who underwent neoadjuvant chemotherapy (NAC). Methods: Texture analysis was performed in a cohort of BC patients who underwent PET/CT at initial staging. Part of these patients then underwent NAC consisting of an association of taxanes and anthracyclines with or without trastuzumab in tumors expressing HER2. For these patients, pCR was assessed using Sataloff classification defined as TA-NA/B. PET/CT images (GE Discovery 690 PET/CT scanner, OSEM reconstruction, voxel size of 2.7 x 2.7 x 3.3 mm3) were obtained 60 to 80 min after injection of 3-3.5 MBq/kg of 18F-FDG when capillary blood glucose was below 11 mMol/L. SUVs (max, mean, peak), 4 histogram-based features and 31 textural indices (TI) were calculated after automatic segmentation of the tumor (40% of SUVmax using LIFEx software). Anova tests with post-hoc analysis using correction for multiple testing were used to compare RF and BC factors (AJCC staging, presence of elevated tumor markers, Ki-67 expression, tumor grade, presence of necrosis, of in situ carcinoma and of an inflammatory stroma) between subtypes: HER2+, Luminal A, Luminal B HER2-, Luminal B HER2+ and Triple-Negative (TN). For each BC subtype, we defined a “radiomic dictionary” including only the RF that were significant different from at least another subtype on post-hoc analysis of Anova tests. Similarly, we defined an “immuno-histo-chemical (IHC) dictionary” based on BC factors. The ability of dictionary features to predict pCR was determined for each BC subtype. Results: A population of 187 patients with a total of 200 breast tumors (>3 mL) was included. HER2+ tumors were excluded because of insufficient recruitment. The “radiomic dictionaries” included from 8 RF (Luminal B HER2+) to 22 RF (Luminal A). Six RF were included in the “radiomic dictionary” of all BC subtypes: SUVmean, SUVmax, SUVpeak, Homogeneity, Dissimilarity and Zone Percentage. The “IHC dictionaries” consisted of 2 BC factors (Luminal A) to 5 BC factors (TN). The usefulness of the subtype-dependent “radiomic dictionaries” and “IHC dictionaries” ictionaries to predict pCR was assessed in a subgroup of 78 patients with 80 tumors (2 patients had multifocal tumors) that underwent NAC: 31 Luminal B HER2-, 16 Luminal B HER2+ and 33 TN tumors. In Luminal B HER2+ tumors, ROC analyses showed that pCR was significantly (p<0.05) associated with 5 RF of the “radiomic dictionary”: SUVmean, SUVmax, LGRE, HGRE and SRLGE and 1 BC factor of the “IHC dictionary”: Tumor grade. In these tumors, the combination of 3 RF (SUVmax, LGRE, and SRLGE) and tumor grade yielded the best performance to predict pCR with a sensitivity/specificity of 100%/83%. By contrast, mixed results were found in TN tumors whereas in Luminal B HER2- no statistically significant association was found. Conclusions: We demonstrated the existence of significant differences in breast tumor metabolic phenotype on 18F-FDG PET/CT and in BC prognostic factors as a function of BC subtype and defined specific “radiomic” and “IHC” dictionaries for each subtype. Combining RF and BC factors from the Luminal B HER2+ dictionaries could help predict pCR.