To detect cancer earlier, we need to predict who will have it in the future. The complex nature of risk prediction has been reinforced by artificial intelligence (AI) tools, but the adoption of AI in medicine has been limited by poor performance in new patient populations and neglect of racial minorities.
Two years ago, a team of scientists from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Jameel Clinic (J-Clinic) demonstrated a deep learning system for predicting cancer risk using only a patient’s mammogram. . The model showed significant promise and even improved inclusion: it was equally accurate for white and black women, which is especially important given that black women are 43 percent more likely to die of breast cancer.
But to integrate image-based risk models into clinical care and make them widely available, the researchers say the models needed algorithmic improvements and large-scale validation at various hospitals to prove their robustness.
To that end, they adapted their new “Mirai” algorithm to capture the unique requirements of risk modeling. Mirai, together, models a patient’s risk at various points in the future and can optionally benefit from clinical risk factors, such as age or family history, if available. The algorithm is also designed to produce consistent predictions in small variations in clinical settings, such as the choice of the mammography machine.
The team trained Mirai on the same data set from over 200,000 Massachusetts General Hospital (MGH) exams from her previous work and validated it on test sets from MGH, the Karolinska Institute in Sweden and the Chang Gung Memorial Hospital in Taiwan. Mirai is already installed at the MGH, and the team’s collaborators are actively working on integrating the model with customer service.
Mirai was significantly more accurate than previous methods in predicting cancer risk and identifying high-risk groups across all three data sets. When comparing high-risk cohorts in the MGH test set, the team found that their model identified almost twice as many future cancer diagnoses compared to the current clinical standard, the Tyrer-Cuzick model. Mirai was equally accurate in patients of different races, age groups and breast density categories in the MGH test set and in different cancer subtypes in the Karolinska test set.
“Enhanced breast cancer risk models allow targeted screening strategies that achieve earlier detection and less damage to screening than existing guidelines,” says Adam Yala, PhD student at CSAIL and lead author of a published article on Mirai this week on Science, Translational Medicine. “Our goal is to make these advances part of the standard of care. We are partnering with doctors at Novant Health in North Carolina, Emory in Georgia, Maccabi in Israel, TecSalud in Mexico, Apollo in India and Barretos in Brazil to further validate the model in diverse populations and study the best way to implement it clinically. ”
How it works
Despite the widespread adoption of breast cancer screening, the researchers say the practice is fraught with controversy: more aggressive screening strategies aim to maximize the benefits of early detection, while less frequent tests aim to reduce false positives, anxiety and costs for those who will never develop breast cancer.
Current clinical guidelines use risk models to determine which patients should be recommended for complementary imaging and MRI scans. Some guidelines use age-only risk models to determine whether, and how often, a woman should be screened; others combine several factors related to age, hormones, genetics and breast density to determine further tests. Despite decades of effort, the accuracy of the risk models used in clinical practice remains modest.
Recently, risk models based on deep learning mammography have shown promising performance. To bring this technology to the clinic, the team identified three innovations that they believe are critical for risk modeling: joint modeling time, the optional use of non-imaging risk factors and methods to ensure consistent performance in environments clinical.
1 time
Inherent in risk modeling is learning from patients with different follow-up periods and assessing risk at different times: this can determine how often they are examined, whether they should have additional imaging tests or even consider preventive treatments.
While it is possible to train separate models to assess risk for each moment, this approach can result in risk assessments that make no sense – such as predicting that a patient is at a higher risk of developing cancer in two years than in five years. To address this, the team designed their model to predict risks at all time points simultaneously, using a tool called the “additive risk layer”.
The additive risk layer works as follows: your network predicts a patient’s risk at one point in time, like five years, as an extension of risk at the previous point, like four years. In doing so, your model can learn from the data with varying amounts of follow-up and then produce self-consistent risk assessments.
2. Non-image risk factors
Although this method focuses primarily on mammograms, the team also wanted to use risk factors with no image, such as age and hormonal factors, if they were available – but did not require them at the time of testing. One approach would be to add these factors as an entry to the model with the image, but this design would prevent most hospitals (such as Karolinska and CGMH), which do not have this infrastructure, from using the model.
In order for Mirai to benefit from risk factors without requiring them, the network provides this information at the time of training and, if it is not there, it can use its own predictive version. Mammograms are rich sources of health information, and many traditional risk factors, such as age and menopausal status, can be easily imaged. As a result of this project, the same model can be used by any clinic globally and, if they have this additional information, they can use it.
3. Consistent performance in clinical settings
To incorporate models of risk of deep learning in clinical guidelines, models must work consistently in different clinical settings, and their predictions cannot be affected by small variations, such as on which machine the mammogram was performed. Even in a single hospital, scientists found that standard training did not produce consistent predictions before and after a change in mammography machines, as the algorithm could learn to rely on different clues specific to the environment. To deviate the model, the team used a contradictory scheme in which the model specifically learns representations of mammography that are invariable to the clinical environment of origin, to produce consistent predictions.
To further test these updates in a variety of clinical settings, scientists evaluated Mirai on new test sets from Karolinska, Sweden, and Chang Gung Memorial Hospital, Taiwan, and found that it achieved consistent performance. The team also analyzed the model’s performance in races, ages, and breast density categories in the MGH test set and in cancer subtypes in the Karolinska data set, and found that the performance was similar across all subgroups.
“African American women continue to develop breast cancer at a younger age, and often at more advanced stages,” said Salewai Oseni, a breast surgeon at Massachusetts General Hospital who was not involved in the work. “This, together with the highest number of triple-negative breast cancer cases in this group, has resulted in increased breast cancer mortality. This study demonstrates the development of a risk model whose prediction is remarkably accurate across races. The opportunity for its clinical use is high. ”
See how Mirai works:
1. The mammography image is passed through something called an “image encoder”.
2. Each image representation, as well as which view it came from, is added to other images from other views to obtain a representation of the entire mammogram.
3. With mammography, a patient’s traditional risk factors are predicted using a Tyrer-Cuzick model (age, weight, hormonal factors). If not available, the predicted values are used.
4. With this information, the additive risk layer predicts the risk of one patient for each year for the next five years.
Improving Mirai
Although the current model does not look at any of the patient’s previous imaging results, changes in images over time contain a wealth of information. In the future, the team plans to create methods that can effectively use a patient’s complete image history.
Similarly, the team notes that the model can be further improved with the use of “tomosynthesis”, an X-ray technique for screening asymptomatic cancer patients. In addition to improving accuracy, additional research is needed to determine how to adapt image-based risk models for different mammography devices with limited data.
“We know that MRI can detect cancer before a mammogram and that early detection improves patient outcomes,” says Yala. “But for patients at low risk for cancer, the risk of false positives may outweigh the benefits. With improved risk models, we can design more subtle risk screening guidelines that offer more sensitive screening, such as MRI, for patients who will develop cancer, to get better results while reducing unnecessary screening and excessive treatment for the rest. ”
“We are excited and humble to ask if this AI system will work for African American populations,” said Judy Gichoya, MD, MS and assistant professor of interventional and computer radiology at Emory University, who was not involved in the work. “We are studying this issue thoroughly and how to detect flaws.”
Yala wrote the article about Mirai alongside MIT research specialist Peter G. Mikhael, radiologist Fredrik Strand of Karolinska University Hospital, Gigin Lin of Chang Gung Memorial Hospital, associate professor Kevin Smith of the KTH Royal Institute of Technology, professor Yung-Liang Wan from Chang Gung University, Leslie Lamb from MGH, Kevin Hughes from MGH, senior author and professor at Harvard Medical School Constance Lehman at MGH, and senior author and professor at MIT Regina Barzilay.
The work was supported by grants from Susan G Komen, the Breast Cancer Research Foundation, Quanta Computing and the MIT Jameel Clinic. It was also supported by the Chang Gung Medical Foundation Grant and the Stockholm Läns Landsting HMT Grant.