Breast cancer is the most frequently diagnosed, life-threatening cancer worldwide and the leading cause of cancer death worldwide in women.
Since the 1970s, breast cancer mortality has been on the decline due to the elimination of hormone replacement therapy, better systemic treatment, and improved screening with mammography.
Screening for breast cancer largely relies on mammography, which is a specialized X-ray image of the breast. An abnormal mammogram shows a soft tissue mass with micro-calcifications, such as the following:
The utility of mammography for breast cancer screening has recently been debated. Mammography screening is limited by the fact that radiologist assessment often varies person to person, with agreements ranging from 62-87% (Spayne et al. Breast J 2012, Sprague et al. Ann Intern Med 2016). As a result, patient mammograms are often reanalyzed between different medical institutions. A recent New England Journal article from Welch et al. argues that screening mammography leads to the over-diagnosis and over-treatment of small tumors that are unlikely to grow and unlikely to lead to mortality.
The large number of false-positive mammograms lead to increased stress, additional medical costs, unnecessary biopsies and treatment. Thus, there is a great need to improve our current screening for breast cancer.
Artificial Intelligence for Mammography Assessment
Computer-aided detection (CAD) to identify distortions in normal breast tissue architecture is on the rise. A recent study by Lehman et al. from the Massachusetts General Hospital demonstrated the potential application of artificial intelligence for mammogram assessment. Full text: https://pubs.rsna.org/doi/10.1148/radiol.2018180694
In this study, Lehman et al. trained a deep convolutional neural network on 58,894 mammograms, showed high agreement of breast density assessment with experienced radiologists (K=0.85), and showed that AI-automated preliminary assessments were accepted as final reads 90% of the time. Even more amazing is that this tool is publicly available for free at http://learningtocure.csail.mit.edu. These free AI programs have tremendous value in resource-poor settings where experienced radiologists are not available and where breast cancer mortality is the largest. The key barrier in a lot of image analyses are access to experts who spent their entire professional lives training their brains to identify pathology. By providing near expert level analyses, these open AI tools have the capability to dramatically improve diagnosis and treatment.
The DREAM Digital Mammography Challenge
In response to efforts to improve breast cancer detection algorithms, IBM and Sage Bionetworks announced the DREAM Digital Mammography challenge.
Funded in part by the Laura and John Arnold foundation, $1.2 million was awarded to teams that could best identify the cancer status in each patient based on digital mammography alone +/- clinical demographic information using 640,000 images from 86,000 women.
More than 120 independent teams submitted to this challenge. The winning teams include Yaroslav Nikulin from Therapixel and Yuanfang Guan from the University of Michigan. These teams created an deep learning algorithms, which accuracy of ~80%, around 5% better than current state of art methods.
Therapixel is a French imaging company that has recently launched Radvise to create an automated first-reads of breast imaging with an fully-secured cloud-based storage system. They also have other products for touch-less medical image navigation in surgery and post-processing image analysis workstations. Website: http://www.therapixel.com/.
Yuanfang Guan is an Associate Professor in the Department of Computational Medicine at the University of Michigan. Since 2013, she has won 10+ DREAM challenges using artificial intelligence to improve image diagnosis. Her team has developed algorithms to detect Parkinson’s disease with mobile devices, sleep apnea using monitors, cancer drug synergy, transcription factor prediction, patient survival and more. Her lab website is here: http://guanlab.ccmb.med.umich.edu/.
Keep an eye out for both of these promising groups!