Cancer Recurrence and Survival Prediction and Evaluation using Machine Learning

• Author / Creator

  Farrokh, Mahtab

• As cancer is the leading global cause of death, an ongoing challenge is predicting an individual's cancer progression accurately, to facilitate personalized treatment planning. Individuals diagnosed with cancer may succumb to the illness or face cancer recurrence post-treatment. The first part of this thesis focuses on predicting prostate cancer recurrence using tissue images. Roughly 30% of men with prostate cancer who undergo radical prostatectomy (RP) will suffer biochemical cancer recurrence (BCR). Unfortunately, no current method can effectively predict which patients will experience BCR after RP. We develop and evaluate PathCLR, a novel semi-supervised method that learns a model that can use tissue images along with clinicopathological features to predict prostate cancer recurrence within five years after RP. We built and evaluated models using two prostate cancer datasets: CPCTR and JHU. PathCLR’s (10-fold cross-validation) F1 score was 0.61 for CPCTR and 0.85 for JHU, which were statistically superior to the best-learned model that relied solely on clinicopathological features. This finding suggests that there is essential predictive information in tissue images at the time of surgery that goes beyond the knowledge obtained from reported clinicopathological features, helping predict the patient’s five-year outcome.The second part of this dissertation focuses on effective survival prediction and evaluation for cancer patients. In the context of deploying individual survival prediction models, a pivotal question emerges: Are we striving to compare survival durations between patients (i.e., ‘Who survives longer between patients A and B?’) or are we endeavoring to estimate a specific patient's survival time (i.e., ‘How long will patient A survive?’), among other scenarios. We address this fundamental inquiry and conduct a comprehensive evaluation of such predictive models. We consider 9 common solid tumors (breast, lung, prostate, etc.) using data from the Surveillance, Epidemiology, and End Results (SEER) Program. We consider several different possible goals of a survival prediction model and connect each goal to a specific evaluation metric. We propose modified versions of the Mean Absolute Error (MAE) measure tailored to address a query about a patient’s expected survival duration. Here, we trained multiple models (including both conventional and advanced machine learning models) on various cancer types and rigorously evaluated those models using the proposed metrics. We demonstrate that a model might be effective for one goal but ineffective for another, and show that we can determine this based on the measure used. Our findings underscore the importance of selecting the evaluation measure that is aligned with the primary objective of a study. This research sets a path for future research that seeks to further refine predictive models for oncological prognostication.

• Subjects / Keywords

  • Machine Learning
  • Survival Analysis
  • Computer Vision
  • Medical Imaging
  • Cancer Study
  • Contrastive learning

• Graduation date

  Spring 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-9j8z-v309

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.