- This event has passed.
Thesis Defence: Training Strategies for Deep Learning-Based Breast Cancer Risk Prediction from Mammography
August 17 at 11:00 am - 3:00 pm
Teymur (Tim) Mammadov, supervised by Drs. Rasika Rajapakshe and Mohamed Shehata, will defend their thesis titled “Training Strategies for Deep Learning-Based Breast Cancer Risk Prediction from Mammography” in partial fulfillment of the requirements for the degree of Master of Science in Computer Science.
An abstract for Tim Mammadov’s thesis is included below.
Defences are open to all members of the campus community as well as the general public. Registration is not required for in-person defences.
Abstract
Breast screening remains the standard of care for early detection of breast cancer. The screening program in BC consistently identifies at least 88% of cancers, however, the abnormal call rate (ACR) remains above the national target. Abnormal mammograms require further procedures, which may or may not confirm presence of breast cancer. High ACR suggests that a number of healthy patients are recalled for additional tests, which causes unnecessary stress and inflicts extra costs. Automated AI (artificial intelligence) systems could assist radiologists in reducing the number of false positives by providing individual risk assessments based on mammography images. Machine learning is already widely used for cancer detection, but the research surrounding deep learning models designed for estimating the risk of future cancers is still limited.
In this thesis, a modern deep learning model was developed using patient mammograms to predict the risk of cancer within five years after a screening exam. This model relies on state-of-the-art (SOTA) architecture and attempts to achieve reliable performance on local data. To that end, a large cohort of patient data from British Columbia (BC) was curated for vision-based learning. It includes 1.8 million images from over 200,000 patients, making it comparable to other large-scale SOTA datasets found in the literature. The model achieves 0.92 [0.89-0.94] and 0.78 [0.76-0.80] ROC AUC for year one and five predictions, compared to 0.84 [0.80-0.88] and 0.76 [0.73-0.78] achieved by pre-trained SOTA in the respective years. These results can be attributed to thorough data curation, two-stage training, and
the informed implementation of the SOTA architecture.
