Abstract
422
Objectives We assessed the ability of textural features to predict recurrence of cervix cancer using pre-treatment 18F-FDG PET images.
Methods 118 patients with cervical cancer treated in Gustave Roussy (GR) between 2005 and 2014 were retrospectively included. All patients underwent a 18F-FDG PET-CT scan in GR before radiotherapy: 77 on a Siemens Biograph scanner (G1 group) and 41 on a GE Discovery scanner (G2 group). PET images were acquired 60.1±4.9 min post-injection. Treatment consisted of a concomitant chemoradiation delivering 45 Gy in 25 fractions of 1.8 Gy to the pelvis +/- the para-aortic area (depending on the results of a primary paraortic surgical staging), followed by a pulse-dose rate image-guided adaptive boost of uterovaginal brachytherapy, aimed at delivering 15 Gy to 90% of the intermediate risk clinical target volume. The primary tumor was delineated using a threshold of 40% SUVmax within a manually drawn volume of interest (VOI). Five conventional indices (SUVmean, SUVmax, SUVpeak in a 1 mL sphere, metabolic volume, TLG) and 6 textural features were calculated in 3D after resampling the VOI SUV between 0 and 20 using 64 gray levels: Homogeneity and Entropy from the Gray-Level Co-occurrence Matrix, Short-Run Emphasis (SRE), Long-Run Emphasis (LRE), Low Gray-level Zone Emphasis (LGZE) and High Gray-level Zone Emphasis (HGZE). ROC analyses were performed in G1 and G2 to study whether the measured parameters were significantly different between relapsing and non-relapsing patients. To determine whether the same threshold values could be used for the two scanners for patient classification, we performed a ROC analysis when merging G1 and G2. Binomial logistic regression including all indices was performed to investigate whether combining all indices could improve patient classification.
Results 27 patients from G1 and 13 from G2 relapsed during the follow-up time (3.3±2.1 years, minimum 1 year). Only Entropy (AUC G1 = 0.709, AUC G2 = 0.698), LGZE (AUC G1 = 0.664, AUC G2 = 0.706), SUVmean (AUC G1 = 0.661, AUC G2 = 0.701) and SUVmax (AUC G1 = 0.664, AUC G2 = 0.695) identified in G1 and G2 the patients who later showed tumor recurrence (p < 0.05). When merging the 2 groups, these indices were still predictive of tumor recurrence with AUC systematically lower (Entropy: 0.670, LGZE: 0.671, SUVmean: 0.662, SUVmax: 0.667) than the best performance achieved by each index in G1 or G2, due to changes in index values between the 2 scanners. With the binomial logistic regression, AUC were significantly higher than using individual indices (AUC G1 = 0.82, AUC G2 = 0.81, AUC G1+G2 = 0.75).
Conclusions Some tumor textural indices measured in baseline FDG PET were as good as SUVs to predict tumor recurrence in 2 patient cohorts scanned with different machines. Although the optimal performance (not achievable in practice) obtained by combining all indices was significantly higher than using a single index, using more biomarkers is required to accurately predict tumor recurrence. Textural and conventional values being different between scanners, the use of a machine-independent threshold for tumor classification is suboptimal.