Room: AAPM ePoster Library
Palliative radiotherapy treatment planning for bone metastases is often based solely on a patient’s simulation-CT scan without recourse to diagnostic CT or PET images. However, using the simulation-CT alone makes accurate detection of metastases regions difficult. Therefore, in this research project, we developed a radiomics-based machine learning (ML) technique to detect lytic bone lesions in simulation-CT images alone.
The simulation-CT-dataset comprised 54 patients with lytic T-spine metastases and 53 control patients with non-metastatic lung cancer. The location of either metastases or healthy bones were labeled with the help of a collaborating radiation oncologist. Regions of interest (ROIs) with various geometric shapes (spherical, cubic, and cylindrical-along-z-axis volumes) were delineated on the images and 104 radiomics features extracted from each ROI using pyradiomics. These radiomic features were used to train and test three ML classifiers: Support Vector Machines (SVM), Random Forest (RF), and Neural Network (NN). The training set comprised 75 patients (38 controls and 37 metastatic), and the remaining 32 were reserved to validate the accuracy, sensitivity, and specificity of each ML method.
The ML algorithm performed best when used with a 5 cm diameter and height cylindrical ROI. For this ROI, NN, RF, and SVM classifiers resulted in 82%, 76%, and 73% accuracy, respectively. The accuracy of the NN model dropped to 76% and 73% when using cubic and spherical ROIs of the same dimensions. The sensitivity and specificity of our best model were 86% and 79%, respectively.
Our radiomics-based ML methods were successful in identifying lytic bone metastases regions in simulation-CT images using a single point based cylindrical ROI. Future work will involve training and validation on much larger datasets, with a goal to develop a tool to assist radiation oncologists to better pinpoint metastatic bone lesions in the simulation-CT images of palliative-intent radiotherapy patients.
Funding Support, Disclosures, and Conflict of Interest: I would like to acknowledge support from the startup grant of Dr. John Kildea at RI_MUHC, RI-MUHC studentship scholarship, and travel grants from McGill Graduate and Postdoctoral Studies.