SBIR Phase I: Exploring bias in Deep learning to extend its use to under-represented populations in breast imaging

Deephealth, Inc., Belmont MA

Investigators

Abstract

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will result from the ability to deploy Artificial Intelligence (AI) to enable rapid and automated detection of breast cancers in mammograms, with the added benefit of greater training of the systems of models on data across a range of demographics, where variability could potentially cause increased errors or inadvertent bias. Breast cancer is the most commonly diagnosed cancer in women and the second leading cause of cancer deaths among women. X-ray mammography screening is an effective tool for reducing mortality among women, but the high volume of mammograms generated every year, combined with the visual challenges of identifying subtle abnormalities in a complex background, makes mammogram analysis a difficult task. This project proposes an AI-based system able to autonomously analyze mammograms to help radiologists in this task, with a broad dataset covering multiple demographics. The proposed solution will save the time of radiologists, allowing the healthcare system to care for women more efficiently, while providing high accuracy and reducing variability of interpretation. The proposed solution may result in a broader adherence to mammography screening programs, marked reduction in breast cancer mortality, and significant cost savings for society. Additionally, this project may lead to improved treatments of other types of cancer. This Small Business Innovation Research (SBIR) Phase I project proposes to develop an algorithm to reduce bias in AI methods for breast imaging in order to extend their use to under-represented demographic groups. Most mammogram databases currently available to train AI models are potentially undertrained for use in under-represented populations because of the limited origin of the test and training data, potentially leading to errors or reduced accuracy when used on other populations. This project proposes a data augmentation procedure to reduce bias by synthetically increasing the number of training examples for under-represented demographics. Firstly, demographic information will be extracted from available databases, and biasing factors will be identified. Then, a new procedure will be established and used to generate highly realistic synthetic mammograms to increase the number of training examples in the original database. The benefits obtained with this procedure will be assessed by training an AI on the original and on the augmented databases and comparing performances in terms of area under the receiving operator characteristic. At the completion of this Phase I project, the augmentation procedure will be ready for testing on under-represented minorities. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.