Prediction of complete pathological response after neoadjuvant chemotherapy in breast cancer using texture analysis: comparison of FDG PET-CT and DCE-MRI.

493 Introduction: The usefulness of neo-adjuvant chemotherapy (NAC) in breast cancer (BC) has been well-established and NAC is now routinely used especially in patients with inoperable tumors or to permit breast conservative surgery in young patients. Yet, this treatment is not devoid of toxicity. It is thus important to define which patients will benefit from NAC as BC patients who achieve complete pathological response (pCR) on the surgical specimen have a better prognosis. FDG PET/CT with SUVmax changes during NAC has shown good results to predict pCR and DCE-MRI (Dynamic Contrast-Enhanced) is being more and more used in this setting (Tateishi et al Radiology 2012). However, the usefulness of texture indices (TI) reflecting tumor heterogeneity to predict pCR has not been fully explored. Here, we compare the performance of TI calculated on FDG PET/CT and on DCE-MRI at initial staging to predict pCR after NAC in BC patients. Methods: Our population was recruited from BC patients who underwent both PET/CT (3-3.5 MBq/kg of 18-FDG) and DCE-MRI (bilateral breast coil and 1.5 T field strength) before NAC associating anthracycline and taxane with the addition of trastuzumab in patients with HER 2+ tumors. Pathological response was assessed on the surgical specimen after NAC according to Sataloff classification with pCR defined as the absence of residual tumor and lymph node involvement. All feature calculation was done using LIFEx software (www.lifexsoft.org). In FDG PET/CT, the tumor was segmented using 40% of SUVmax and voxel intensities were resampled using 64 discrete values (DV) between 0 and 20 SUV units (bin width: 0.3). In DCE-MRI acquired 2 minutes after Gadolinium IV injection, the whole tumor volume was delineated manually and 200 DV were used between 0 and 20000 units (bin width: 100) to calculate the TI. A total of 31 TI including 6 robust TI (homogeneity, Entropy, SRE, LRE, LGZE and HGZE) and 4 Histogram Based Index (HBI: SkewnessH, kurtosisH, EntropyH and EnergyH) were derived, and SUVs (mean, max and peak) were also calculated from PET images. All these features were assessed for their performance to discriminate between tumors that showed pCR on the surgical specimen compared to those that had residual tumor (RT) using multivariate and ROC analyses. Results: Thirty-nine BC patients with a total of 41 tumors that underwent NAC before surgery were recruited. When we look at the whole cohort, there were significant differences on FDG PET/CT between tumors which yielded pCR compared to RT for 6 TI including 1 robust TI (HGZE). Only one TI (SZE) was significantly different between the pCR and RT groups on DCE-MRI. ROC analyses yielded AUC values >0.7 only for TLG and LRHGE and for SZE in DCE-MRI, but not for SUVs in FDG PET/CT. When looking at the 33 HER 2- tumors only (different NAC regimen: no trastuzumab), there were significant differences for 15 TI including 4 robust TI (homogeneity, entropy, SRE and HGZE) on FDG PET/CT but none on DCE-MRI. Similarly on ROC analysis in HER 2- tumors, 24 TI including the 6 robust TI showed AUC values >0.7 to predict pCR on FDG PET/CT but only one TI (Contrast) on DCE-MRI. When separating the tumors as a function of their BC subtypes, there were significant differences for 4 TI including 2 robust TI (LGZE and HGZE) between tumors that yielded pCR compared to RT for basal-like subtype but none for the Luminal B (HER 2- and 2+) tumors. Conversely, there were no significant differences found for any of the BC subtypes on DCE-MRI. Conclusion: Texture analysis better predicts pCR in BC on FDG PET/CT than on DCE-MRI especially when NAC regimen depending on the HER 2 status is taken into account. Whether combining FDG PET/CT TI with DCE-MRI TI in multivariate radiomic models enhances the predictive power of FDG PET/CT TI remains to be investigated. Also, the calculation of TI in DCE-MRI might have to be optimized (for instance using different DV or bin width) to get more significant results from the MR images.