Author: Issam M. El Naqa, Kurukulasuriya Ruwani Fernando, Himani Himani, Vivek Kumar, Arun Oinam, Manju Sharma 👨🔬
Affiliation: Panjab University, Moffitt Cancer Center, H. Lee Moffitt Cancer Center, Post Graduate Institute of Medical Sciences, University of California San Francisco 🌍
Purpose: To investigate the utility of Magnetic Resonance Imaging (MRI)-based radiomics for predicting tumor response and adverse effects, specifically gastrointestinal (GI) toxicity, in cervical cancer patients undergone radiotherapy.
Methods: A retrospective analysis was conducted on 45 cervical carcinoma patients treated with radiation therapy between October 2010 and January 2015. The patient follow up was done for 5 yearsT2-weighted MRI scans were used to extract 851 radiomic features, which were pre-processed to ensure relevance and avoid redundancy. Feature selection was performed using the Pearson Correlation Coefficient (p > 0.95) and Random Forest Classifier to identify the most predictive features. To address the inherent class imbalance in the dataset, Synthetic Minority Oversampling Technique (SMOTE) and other techniques were applied. Predictive models, including logistic regression, random forest, decision trees, and gradient boosting classifiers, were constructed and assessed using cross-validation to predict CTCAE toxicity grades. Model evaluation metrics included accuracy, precision, recall, F1-score, confusion matrix, and the area under the receiver operating characteristic curve (ROC AUC).
Results: The feature selection process effectively identified a subset of radiomic features with significant predictive capability. Among the models tested, the random forest classifier achieved the highest performance, with a mean accuracy of 73.3±1.5%, precision of 72.4±2.2%, recall of 69.1±2.1%, F1-score of 65.5±1.9%, and an ROC AUC of 0.60±0.25. The gradient boosting classifier also demonstrated competitive results, underscoring its potential in radiomics-based predictions.
Conclusion: Preliminary results demonstrates the feasibility of leveraging radiomics in MRI for predicting tumor response and radiation-induced GI toxicity in cervical cancer patients. The findings highlight the potential of integrating advanced feature selection and predictive modeling techniques to enhance personalized treatment planning and improve patient outcomes in radiotherapy. Further validation in larger cohorts and clinical trials is necessary to confirm the clinical utility of these radiomic biomarkers in predicting radiotherapy-related toxicity.