3391 - A Machine Learning Model Based on Computed Tomography Radiomics for the Prediction of the Response to SBRT in Lung Metastasis from Colorectal Cancer
Presenter(s)
D. Caivano1,2, F. Palmeri3, A. Del Gaudio3, P. Bonome4, D. Pezzulla4, C. Scaringi1, G. Apicella1, D. Musio1, D. Caruso3, and M. F. Osti2,5; 1Department of Radiation Oncology, San Giovanni Addolorata Hospital, Rome, Italy, 2PhD School in Traslational Medicine and Oncology, Department of Medical and Surgical Sciences and Translational Medicine, Faculty of Medicine and Psycology, Sapienza University of Rome, Rome, Italy, 3Sapienza - University of Rome, Department of Surgical and Medical Sciences and Translational Medicine - Radiology Unit - Sant'Andrea University Hospital, Rome, Italy, 4Radiation Oncology Unit, Responsible Research Hospital, Campobasso, Italy, 5Radiotherapy Department, St. Andrea Hospital, Sapienza University of Rome, Rome, Italy
Purpose/Objective(s): The aim of this work was to develop a machine learning model based on computed tomography images to stratify response to SBRT for lung metastases (LM) from colorectal cancer (CRC).
Materials/Methods: We collected 88 CT scans, each with one VOI. Of these, 24 (27.3%) belonged to the progressive class and 64 (72.7%) belonged to the non-progressive class, according to response to treatment SBRT. This image set was used for the training, validation, and testing of five machine-learning models. Moreover, we also collected 16 samples, to form an external testing set to test the best model obtained, composed of 5 samples (31.3%) belonging to class "Progressive" and 11 samples (68.8%) belonging to class "Non progressive". Statistical analysis was performed using tools of the Trace4Research platform. We described the distribution of each feature in the "progressive" and "non-progressive" classes using medians and 95% CIs. A non-parametric test was performed on each relevant radiomic predictor to test their significance in discriminating between the 'progressive' and 'non-progressive' classes. P-values were adjusted for multiple comparisons.
Results: From each segmented VOI of each image considered in this study, IBSI-compliant radiomic features and features based on deep learning were computed, for a total of 3738 features. Of these, 6 were finally chosen. For the classification task of interest (24 images from class "Progressive" vs. 64 images from class "Non progressive"), these features were used for training, validation, and testing (nested 10-fold cross validation) of the five different models of machine-learning classifiers considered in this work. Based on ROC-AUC, the model of the Support Vector Machine classifiers emerged as the best model for the task of interest ("Progressive" vs. "Non-progressive") and was therefore tested on the external test set (5 samples (31.3%) belonging to the "Progressive" class and 11 samples (68.8%) belonging to the "Non-progressive" class). The 6 radiomic predictors of interest are ranked in terms of their statistical significance. The median values of each characteristic, the 95% CIs and the results of the univariate Wilcoxon rank sum test are also reported, together with the adjusted p-values.
Conclusion: Research in this area aims to identify potential applications of radiomics as a predictor of local response to SBRT for lung lesions. In the future, we hope to collect even more data to provide the best answers for predicting treatment outcomes. This will help us to provide increasingly personalized medicine.