3630 - Magnetic Resonance Imaging-Based Radiomics Analysis for the Prediction of Overall Survival in Locally Advanced Cervical Cancer Patients
Presenter(s)
M. C. Hormazabal1, I. Liedtke1, G. Lazcano Alvarez2,3, A. Molina4, J. Giusti-Bilz4, F. Olivares5, F. Pérez Peña4, D. Folch4, T. W. Martin2, and J. A. solis Campos1,3; 1Hospital Carlos Van Buren, Valparaíso, Chile, 2Hospital Carlos Van Buren, Valparaiso, Chile, 3Universidad de Valparaíso, Valparaíso, Chile, 4Universidad de Valparaiso, Valparaiso, Chile, 5Universidad de Valparaiso, valparaiso, Chile
Purpose/Objective(s): Radiomics is an approach that involves extracting quantitative features from medical images to enhance the development of models that could support clinical decision-making. The purpose of this study is to develop a predictive model based on MRI-based radiomic features for the 3-year overall survival (3y-OS) in cervical cancer patients treated with concurrent chemoradiotherapy and MRI-based image-guided adaptive brachytherapy (IGABT).
Materials/Methods: A retrospective cohort including locally advanced cervical cancer patients treated at a single center between 2019-2022 was identified. Treatment included concurrent chemoradiotherapy with weekly cisplatin and MRI-based IGABT. T2 weighted magnetic resonance Images (MRIs) acquired before treatment were used to define the Intermediate-risk clinical target (IR CTV). CTV-IR-Radiomic features were extracted from the sagittal MRIs acquired before treatment using open source software. Features were extracted from the original unfiltered images and from a set of eight filtered images, by selecting the following image type parameter options: exponential, gradient, local binary pattern 2D and 3D, Laplacian of gaussian, logarithm, square, square root and wavelet. Radiomic features consist of 7 classes: shape, first order statistics, glcm, gldm, glrlm, glszm and ngtdm. 1967 initial features were extracted. Additional 14 demographic and clinical features were considered as initial input features prior to the feature selection process.
Dimensionality reduction involved:
- Removing highly correlated features (Pearson> 0.99).
- Feature selection by feature importance analysis via iterative Random Forest algorithm.
Multiple data sets containing different numbers of features were used to train different supervised classification algorithms. The model's performance was assessed using the AUC metric with k-fold cross-validation technique. 3y-OS was the endpoint of the study.
Results: 103 patients were included. Median age was 47 years (IQR 37-58). Out of the total, 79 patients achieved the 3y-OS endpoint and 24 patients did not. A classification model using the Support Vector Classification algorithm was developed. No clinical or demographic data were selected by the feature selection process. The model considered 47 radiomic features resulting in an average 5-fold cross-validation AUC of 0.833 (SD 0.032).
Conclusion: A radiomic model, utilizing only radiomic features, was developed to predict 3y-OS in patients treated with concurrent chemoradiotherapy and IGABT. While clinical and demographic variables were available before the feature selection process, only radiomic features were selected. Clinical and demographic data did not contribute to enhancing the model’s predictive power.