Investigating multi-radiomic models for enhancing prediction power of cervical cancer treatment outcomes. (February 2018)

• Record Type: Journal Article
• Authors: Altazi, Baderaldeen A.; Fernandez, Daniel C.; Zhang, Geoffrey G.; Hawkins, Samuel; Naqvi, Syeda M.; Kim, Youngchul; Hunt, Dylan; Latifi, Kujtim; Biagioli, Matthew; Venkat, Puja; Moros, Eduardo G.
• Is Part Of: Physica medica. Volume 46(2018)
• Page Start: 180
• Page End: 188
• Publication Date: 2018-02
• Subjects: Positron emission tomography -- Radiomics -- Tumor uptake -- Cervical cancer; Medical physics -- Periodicals; Biophysics -- Periodicals; Biophysics -- Periodicals; Imagerie médicale -- Périodiques; Radiothérapie -- Périodiques; Rayons X -- Sécurité -- Mesures -- Périodiques; Physique -- Périodiques; Médecine -- Périodiques; 610.153
• DOI: 10.1016/j.ejmp.2017.10.009
• Languages: English
• ISSNs: 1120-1797

Abstract

Graphical abstract: Highlights: Multi-radiomic modeling, enhanced the predictive power for treatment outcomes. Radiomic features shown to be a better outcome predictor than SUV measurements. The metabolic tumor volume is highly predictive for both LRR and DM. Our predictive models were selected using sequential backward selection method as part of LOOCV within the training set. The trained models were re-evaluated on an independent test set. Abstract: Quantitative image features, also known as radiomic features, have shown potential for predicting treatment outcomes in several body sites. We quantitatively analyzed 18 Fluorine–fluorodeoxyglucose ( 18 F-FDG) Positron Emission Tomography (PET) uptake heterogeneity in the Metabolic Tumor Volume (MTV) of eighty cervical cancer patients to investigate the predictive performance of radiomic features for two treatment outcomes: the development of distant metastases (DM) and loco-regional recurrent disease (LRR). We aimed to fit the highest predictive features in multiple logistic regression models (MLRs). To generate such models, we applied backward feature selection method as part of Leave-One-Out Cross Validation (LOOCV) within a training set consisting of 70% of the original patient cohort. The trained MLRs were tested on an independent set consisted of 30% of the original cohort. We evaluated the performance of the final models using the Area under the Receiver Operator Characteristic Curve (AUC). Accordingly, six models demonstrated superior predictive performance for both outcomes (four for DM and two for LRR) when compared to both univariate-radiomic feature models and Standard Uptake Value (SUV) measurements. This demonstrated approach suggests that the ability of the pre-radiochemotherapy PET radiomics to stratify patient risk for DM and LRR could potentially guide management decisions such as adjuvant systemic therapy or radiation dose escalation.