G. Ramasamy1, G. Kasymjanova2, J. Agulnik2, J. M. G. Tsui3,4, and T. M. Muanza5,6; 1McGill University, Montreal, QC, Canada, 2Peter Brojde Lung Cancer Centre, Jewish General Hospital, Montreal, QC, Canada, 3Cedars Cancer Centre, Department of Radiation Oncology, McGill University Health Centre, Montreal, QC, Canada, 4Gerald Bronfman Department of Oncology, McGill University, Montreal, QC, Canada, 5Faculty of Medicine, Experimental Medicine Department, McGill University, Montreal, QC, Canada, 6Radiation Oncology Department, Jewish General Hospital, McGill University, Montreal, QC, Canada
Purpose/Objective(s): Medical images contain a wealth of pathophysiology information that can be extracted and used for clinical decision support systems. Indeed, radiomics features such as skewness, root mean square [Ben Man Fei Cheung et al. 2021], homogeneity and long-run-high-gray-level-emphasis [Kyle J Lafata et al. 2019] have been shown to predict local response for SBRT-treated NSCLC tumors. This guides personalized medicine and we wish to apply it to our cohort. That said, radiomics features are not stable [Mitchell Chen et al. 2023]. Thus, we attempt to validate for use in our institution. This study investigates the radiomic features of images and clinical parameters obtained from early-stage and oligometastatic non-small cell lung cancer (NSCLC) patients who underwent stereotactic body radiation therapy (SBRT) to predict local control. Materials/
Methods: A single-institution retrospective review of patients’ medical records (n = 98 patients; median age = 76 yrs; male/female ratio = 46/52; 116 lesions) treated with SBRT from 2009 to 2022 was conducted. Radiomics features (107 features) extracted from CT planning scans with an open source software, along with patient clinical data were used. Response to SBRT was analyzed from follow-up scans. Response was defined as per RECIST criteria. Complete response (CR) or partial response (PR) lesions were responders (33 lesions) and stable disease (SD) or progressive disease (PD) were non-responders (83 lesions). Adaptive synthetic (ADASYN) sampling corrected the imbalance in responses. Classification models, which included support vector machines (SVM) with linear or radial basis function (RBF) kernels, random forests, adaptive boosting (AdaBoost) and multi-layer perceptron (MLP), were used. Models were trained using a 5-fold cross-validation scheme. Their performances were measured with the areas under the curve (AUC) of receiver operating characteristic (ROC) plots on the validation folds. Using permutation feature importance, predictive biomarkers were identified. Results: The best model used a MLP classifier and had an AUC of 0.94+/-0.05. Treatment site, performance status, along with radiomic features such as first-order root mean squared intensity and GLSZM gray level non uniformity emerged as the most predictive. Other less but still predictive features included: first order skewness, GLCM joint entropy and cluster shade, GLSZM small-area-low-gray-level-emphasis and size-zone-non-uniformity-normalized, GLDM small-dependence-low-gray-level-emphasis and small-dependence-emphasis. Conclusion: Consistent with previous research, root mean squared intensity and skewness were found to be predictive biomarkers., In contrast, homogeneity and long-run-high-gray-level-emphasis did not emerge as biomarkers for local response prediction. Hence why in-house radiomics feature validation is crucial. This demonstrates the instability of certain predictive radiomic features when applied to external cohorts.
