Application of Machine Learning in Prostate Cancer External Beam Radiation Therapy to Improve Biochemical Failure-Free Survival Prediction

Simple item page
dc.contributor.advisorSmith, Wendy
dc.contributor.advisorKirkby, Charles
dc.contributor.advisorBrown, Jo-Anne;
dc.contributor.authorSun, Lingyue
dc.contributor.committeememberQuon, Harvey
dc.contributor.committeememberMcGeachy, Philip
dc.dateFall Convocation
dc.date.accessioned2023-05-11T05:12:23Z
dc.date.embargolift2023-08-24
dc.date.issued2022-08-24
dc.description.abstractCurative external beam radiation therapy (EBRT) is a common treatment option for localized prostate cancer. The reported biochemical failure-free survival (BFFS) for prostate cancer patients post-EBRT varies significantly. The current, widely implemented clinical models that are used to predict treatment outcome for these patients have limited reliability. This thesis aims to examine the possible sources of influence on BFFS and improve EBRT treatment outcome prediction in terms of BFFS in particular using machine learning (ML) algorithms.Radiotherapy clinical trial protocols are adopted by different institutions due to their promising outcome(s) and/or superior efficiency. In this thesis, we first quantified the plan quality differences across four institutions in Alberta for PROstate Fractionated Irradiation Trial (PROFIT) plans. Despite the common guiding protocol, we observed statistically significant differences in dosimetric parameters between institutions; however, outcome modeling suggested such differences were of minimal clinical consequence.There is a need for more detailed understanding of the relationships between EBRT treatment planning and treatment delivery-related features and tumor control outcomes. We investigated the EBRT-related features that may be prognostic of BFFS using a random survival forest model. We found that the clinical target volume D99, pelvic irradiation, IGRT frequency, and planning target volume V98 were prognostic in addition to a set of tumor features.We then investigated whether ML-based models that incorporated additional EBRT treatment planning and delivery related features could perform better than the clinical models that were exclusively based on patient demographic and tumor features. We found that the two ML-based models outperformed the two clinical models on both the training and validation datasets, although all models’ performances deteriorated against the validation dataset.To further improve BFFS prediction, we investigated whether adding dosiomic features, which contains spatial information about the planned 3D dose distribution, could enhance the model performance. We did find improvement, and such improvement although relatively small was maintained when the model was validated with patients from other institutions. Additionally, we evaluated the impact of dose calculation variations on dosiomic features and found that most dosiomic features were stable against variations in dose calculation algorithms, versions, and dose grid.
dc.identifier.citationSun, L. (2022). Application of Machine Learning in Prostate Cancer External Beam Radiation Therapy to Improve Biochemical Failure-Free Survival Prediction (Doctoral thesis). University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca .
dc.identifier.urihttp://hdl.handle.net/1880/116387
dc.identifier.urihttps://dx.doi.org/10.11575/PRISM/dspace/41231
dc.language.isoEnglish
dc.publisher.facultyScience
dc.subject.classificationPhysics--Radiation
dc.subject.classificationHealth Sciences--Oncology
dc.subject.classificationArtificial Intelligence
dc.titleApplication of Machine Learning in Prostate Cancer External Beam Radiation Therapy to Improve Biochemical Failure-Free Survival Prediction
dc.typedoctoral thesis
thesis.degree.disciplinePhysics & Astronomy
thesis.degree.grantorUniversity of Calgary
thesis.degree.nameDoctor of Philosophy (PhD)
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
ucalgary_2022_sun_lingyue.pdf
Size:
8.5 MB
Format:
Adobe Portable Document Format
Download
Collections
Open Theses and Dissertations