Electronic Thesis and Dissertation Repository

Thesis Format

Integrated Article

Degree

Master of Science

Program

Medical Biophysics

Supervisor

Mattonen, Sarah A.

2nd Supervisor

Nguyen, Timothy K.

Affiliation

London Health Sciences Centre

Co-Supervisor

Abstract

Patients diagnosed with non-spine bone metastases benefit from stereotactic body radiotherapy (SBRT), however, 5-20% of patients still develop local progression. We aim to create a machine learning model integrating clinical, radiotherapy dose, and radiomic features to predict which patients will experience local progression following SBRT. We analyzed 179 treated metastatic lesions in 130 patients. Radiomic features were extracted from radiotherapy target volumes. Lesions were separated by patient into an approximate 70/30 training (n = 119) and testing (n = 60) split. Feature selection was performed, and machine learning models were created using dose, clinical, and radiomics features. The clinical model achieved a testing AUC of 0.61 [95% CI: 0.47 – 0.74] and the combined model had a testing AUC of 0.77 [95% CI: 0.64 – 0.89]. This preliminary study demonstrates the potential utility for a radiomics-based machine learning model for the prediction of local progression in non-spine bone metastases following SBRT.

Summary for Lay Audience

Metastases occurs when cancer cells spread from the primary tumour volume to a secondary region within the body. Bone metastases occurs when these cancer cells create a secondary tumour in the bone. Patients with bone metastases are treated with high-dose radiotherapy delivered directly to the tumour volume, which provides minimal dose to the surrounding normal tissues. This high dose radiotherapy provides patients with improved outcomes, such as decreased rates of the cancer returning after treatment, as well as increased survival. However, up to 20% of patients will have the cancer return at the same location after treatment. Medical imaging, such as computed tomography (CT), is used for planning the radiation treatment for these patients. Features from these medical images, such as texture, which a clinician might not see with their eye, can be collected and used to provide further insight into the aggressiveness of the cancer. Clinicians are currently unable to determine prior to treatment which patients will have their cancer return at the treated site. In this study, we aim to use imaging features along with information from the patient, tumour, and radiation treatment plan to predict which patients will have their cancer return after treatment. We used this information to create machine learning software models and found that adding imaging features to patient and tumour information improved the ability to predict cancer returning after treatment. This leads us to believe that imaging information may provide insight into tumour aggressiveness and help clinicians identify patients at a higher risk of their cancer returning after treatment. These models can be used to identify before treatment which patients may benefit from more aggressive treatment, such as a higher dose of radiation.

Download

Share

COinS