Towards Personalized Screening: An AI Framework for Individual Breast Cancer Risk Assessment from MRI

The project aims to develop a deep learning framework for individualized breast cancer risk prediction from MRI, addressing the limitations of current “one-size-fits-all” screening procedures. The project is structured into three objectives:

1. The first objective involves implementing a deep learning framework to train a model on a large, unlabeled dataset of breast MRI volumes from Sunnybrook. This process will focus on making sure that the model is able to learn robust features directly from the data.

2. The utility of this model will be quantitatively assessed by fine-tuning it for specific downstream tasks, namely breast lesion and breast tissue segmentation. The performance will be benchmarked against a model trained from scratch, with our metrics being segmentation accuracy, which will validate the quality of our pre-trained model.

3. The final and primary objective is to leverage the pre-trained model for risk prediction. It will be used to train a predictive model that predicts the probability of breast cancer being developed. The core hypothesis, based on prior work in mammography, is that a high prediction score in a scan from a woman, without a current cancer diagnosis, correlates with a higher future risk.

Faculty Supervisor:

Anne Martel

Student:

Partner:

Eindhoven University of Technology

Discipline:

Engineering

Sector:

Education

University:

University of Toronto

Program: