Description:
Risk assessment is a key tool for personalized screening and early detection of breast cancer. In recent work1, the MIT (Massachusetts Institute of Technology) team developed deep learning models that assess cancer risk from full-resolution screening mammograms and found them to be significantly more accurate than the Tyrer-Cuzick model, identified in the top risk decile by the Hybrid DL model were at 3.8x higher risk than the average woman compared to 1.9x by the Tyrer-Cuzick model. Since this publication, the MIT team has significantly improved the model and extended it to jointly predict the risk of future cancer at one to five years away from the mammogram2. Moreover, they have also developed a new model for risk-based screening recommendations and found it to outperform traditional age-based screening guidelines in retrospective simulations.
Purpose and objectives:
The primary aim of this study is the measure the discriminative performance of the developed risk models on the Ontario Breast Screening Program (OBSP) screening population.
Our secondary aim is to quantify the performance of learned personalized screening recommendations in retrospective analysis. This study will be a retrospective chart review analyzing the charts and mammograms of any female patient who had a screening mammogram from 1/1/2012 to 12/31/2016.
Hypothesis:
Women with the top risk decile by the Hybrid DL model will have a higher cancer detection rate than women in the lower risk decile, thus better predicting breast cancer development and risk assessment than currently exists in the OBSP.
Specific Objective(s):
We will evaluate all women diagnosed with breast cancer within the OBSP program within 2012-2016 to determine how well the Hybrid DL model performed in prediction of breast cancer.
Brief Methods: Analyses of results
To measure the accuracy of each risk model, we will measure the concordance index of the predictions with the provided outcomes on the retrospective data as well at the ROC AUCs at one to five years from the time of the mammogram. To measure the simulated performance of risk-based screening, we will measure relative screening volume under simulated screening guidelines as well as shifts in diagnosis time. We will assess confidence intervals using a clustered-bootstrap procedure. To facilitate this work, we will collect the risk assessment and outcomes for each anonymized patient ID, and this data will be sent to MIT to complete the analysis. Note, no images or identifiable information will be sent to MIT.
MIT will analyze the data as follows:
- Data Collection and Computer Setup
Hampton Park TOH team will prepare the dataset as well as a computing environment. - Analysis of Risk models on Screening Cohort
Using on premise computing at Hampton Park TOH and a virtual private network, the MIT team will process data and compute risk assessments. Given the risk assessments and outcomes, the MIT team will evaluate the performance of different image based risk models. - Analysis of Screening guidelines on Screening Cohort
MIT team will compute simulated impact of risk-based screening for different risk models on screening cohort.
Student skills required:
The student will be required to review patients in EPIC, have Excel and Word skills, be able to review using PACS, and collaborate with the research staff to complete tasks. Prior research collaboration and skills would be an asset.
Potential skills developed:
The student would learn about the imaging for cancer detection, the performance indicators for high quality work in breast imaging, the importance of Artificial intelligence in Radiology, as well as joining a collaborative team of researchers dedicated to early detection of breast cancer.
Timeline:
The project would begin May 1, 2024, and completed by August 31, 2024, with the possibility of data analysis continuing into the fall.
Benefit to researcher/professor:
I am interested in mentoring a medical student who is interested in research in Radiology and Breast Imaging, and to collaborate with them on a project that may improve the quality of breast cancer risk assessment and allow greater access to early detection of breast cancer.
Matching funding:
No
References:
- Yala A, Lehman C, Schuster T, Portnoi T, Barzilay R. A Deep Learning Mammography-based Model for Improved Breast Cancer Risk Prediction. Radiology. 2019;292:60-66. doi: 10.1148/radiol.2019182716
- Yala A, Mikhael PG, Strand F, Lin G, Satuluru S, Kim T, Banerjee I, Gichoya J, Trivedi H, Lehman CD, et al. Multi-Institutional Validation of a Mammography-Based Breast Cancer Risk Model. J Clin Oncol. 2022;40:1732-1740. doi: 10.1200/jco.21.01337