ESTRO 2021 Abstract Book


ESTRO 2021

Table 1: ROC AUC of models

ROC AUC (95% CI) Training (n=200)

Validation (n=104)

Clinical alone

0.73 (0.66-0.80)

0.62 (0.51-0.74)

Radiomics alone

0.68 (0.61-0.75)

0.66 (0.56-0.77)

Clinical and Radiomics


0.68 (0.57-0.79)

Conclusion The predictive characteristics of clinical variables, with or without radiomics, are modest. There is a need to develop new approaches to optimise neoadjuvant strategies for enhancing cCR.

PH-0106 FDG-PET features help predict distant metastases in oropharyngeal cancer patients with definitive RT P. Brodin 1 , J. Lubin 2 , J. Eichler 2 , C. Velten 2 , S. Zhu 3 , S. Saha 3 , W. Tomé 1 , C. Guha 1 , S. Kalnicki 2 , R. Kabarriti 1 , M. Garg 2 1 Albert Einstein College of Medicine and Montefiore Medical Center, Radiation Oncology, Bronx, USA; 2 Montefiore Medical Center, Radiation Oncology, Bronx, USA; 3 Albert Einstein College of Medicine, Radiation Oncology, Bronx, USA Purpose or Objective To evaluate whether primary tumor imaging features from pretreatment positron emission tomography (PET) scans can predict progression-free survival (PFS) and risk of distant metastases (DM), for risk-stratification of patients treated with definitive radiation therapy (RT) and chemotherapy for oropharyngeal cancer. Materials and Methods Patients with oropharyngeal cancer treated from 2005 to 2018 with definitive RT at our institution and received a pre-treatment PET/CT scan were identified. Demographics and clinical outcomes data were collected. Patients were excluded if there was no PET avidity, their primary site of disease was poorly visualized or they received initial surgical resection. Primary tumor volume was delineated on pretreatment PET/CT using a gradient-based segmentation and used for subsequent image feature extraction. Multiple radiomics image features were computed, along with metabolic tumor volume (MTV), SUV mean , SUV max and total lesion glycolysis (TLG). Image features were extracted following isotropic resampling and z-score normalization. Features with zero variance and strong correlation to tumor volume, stage, HPV p16-status, age or smoking were excluded. A random forest classification model with LASSO regularization was used to select image features associated with PFS. Kaplan-Meier methods and receiver operating characteristics (ROC) with 5-fold cross-validation were used to analyze prediction performance. Results 120 patients were identified, 6 excluded due to less than 3 months follow-up. With a median follow up of 39.6 months (range: 3.1 – 137.5) for the remaining 114 patients, 14 had local recurrence, 22 had DM and 38 patients died. The 2-year local control, distant control and PFS were 88.9%, 84.2% and 69.7%, respectively. Three wavelet image features were identified in the random forest model. Table 1 shows that clinical features (HPV p16, stage, age and smoking) provide good predictive performance for PFS with ROC AUC>0.70, and addition of wavelet features improved performance, while addition of simple SUV metrics did not. HPV p16 positivity and age>65y were strong predictors of PFS with multivariable HR=0.30, p=0.001 and HR=3.4, p<0.001, respectively. A radiomics risk signature dichotomizing patients into high and low risk was found to be a strong predictor of DM with a multivariable HR=21.8, p=0.004. This was especially seen for patients with HPV p16 negative disease with a 2-year DM-free survival of 87.4% for low vs. 52.9% for high radiomics risk score (p=0.016).

Made with FlippingBook Learn more on our blog