The Future Brain
By Cami Rosso
New deep learning model by MIT CSAIL predicts breast cancer risk.
Posted May 23, 2019
A few weeks ago, a new type of artificial intelligence (AI) deep learning model was unveiled. Unlike existing models, this new deep learning model can predict future breast cancer risk up to five years in advance with greater accuracy than what is currently used in clinical practice today—regardless of the patient’s racial background. Researchers from Massachusetts Institute of Technology (MIT) and Massachusetts General Hospital have filed patents on the new deep learning model and published their study on May 7, 2019, in Radiology, a peer-reviewed journal.
“We hypothesize that there are subtle but informative cues on mammograms that may not be discernible by humans or simple volume-of-density measurements, and deep learning (DL) can leverage these cues to yield improved risk models,” wrote the researchers. “Therefore, we developed a DL model that operates over a full-field mammographic image to assess a patient’s future breast cancer risk.”
Early cancer detection improves patient outcomes. The percentage of people who live beyond five years after a diagnosis of breast cancer varies by stage—the later the detection, the lower the chances of surviving beyond five years of the diagnosis. Breast cancer survival rates are 98-100 percent for stage I, 90-99 percent for stage II, and 66-98 percent for stage III, according to figures adapted by Susan G. Komen®, a U.S. breast cancer nonprofit organization, from a study published in JAMA Oncology in 2018 by researchers from The University of Texas MD Anderson Cancer Center and the Cancer Prevention Institute of California.
The researchers in this study set out to create a breast cancer model that can assess risk equally well for all patients, regardless of the patient’s racial heritage. The team consists of Adam Yala, the study’s lead author and MIT CSAIL Ph.D. candidate, along with co-authors MIT Professor Regina Barzilay, Harvard Professor Constance Lehman, MIT CSAIL Ph.D. student Tal Schuster, and Tally Portnoi, MIT MEng.
Article Continues After Advertisement
Existing AI models for breast cancer detection are often less accurate in screening patients from minority racial backgrounds, such as African Americans. There are many areas where machine learning bias may occur when developing artificial intelligence models. Bias can be introduced in the data structure, collection, sources, and amount. Bias can also be in the algorithm itself via the weights assigned to the data points and the absence or inclusion of indicators. The machine learning model can be also be influenced by the biases of the people creating and managing the model, depending on how the system is managed.
In the study, three models using artificial intelligence to assess breast cancer risk within five years were developed: 1) a risk-factor-based logistic regression model (RF-LR) that used traditional risk factors, 2) an image-only deep learning model that only used mammograms, which was a deep convolutional neural network (ResNet18) with PyTorch, and 3) a hybrid deep learning model that used both a combination of traditional risk factors and mammograms.
The researchers compared the results from the models to the traditional breast cancer risk model that included breast density, the Tyrer-Cuzick model (version 8). The researchers found that the hybrid deep learning model performed the best overall out of all three models. Specifically, the hybrid deep learning model placed 31 percent of breast cancer patients in the highest risk category versus the Tyrer-Cuzick model (version 8), which only placed 18 percent.
The image-only deep learning model that was based solely on mammograms also performed better than the established Tyrer-Cuzick model (version 8). Another benefit of the image-only deep learning model is that it could provide accurate risk assessment even in the absence of traditional risk factor data, which is helpful in cases where the patient’s family history of cancer is unknown.
“These results support the hypothesis that mammography contains informative indicators of risk not captured by traditional risk factors, and DL models can deduce these patterns from the data,” wrote the researchers. “These models have the potential to replace conventional risk prediction models.”
This new deep learning model enables early breast cancer detection, regardless of the patient’s race—providing hope for a better future for all patients.
Copyright © 2019 Cami Rosso All rights reserved.
Follow Us on Ginnyent.net By Clicking on The Button Below The Page.