To get cancer earlier, we need to predict who will suffer 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 the poor performance of new patient populations and neglect of racial minorities.
Two years ago, a team of scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Jameel Clinic (Clinic J) 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 both white and black women, which is particularly important given that black women are 43% more likely to die from the cause of breast cancer.
But to integrate imaging risk models into clinical care and make them widely available, researchers say the models needed both algorithmic improvements and large-scale validation in several hospitals to demonstrate their robustness.
To this end, they have adapted their new “Mirai” algorithm to capture the unique requirements of risk modeling. Mirai together models a patient’s risk in several future times and may optionally benefit from clinical risk factors, such as age or family history, if available. The algorithm is also designed to produce predictions that are consistent between minor variations in clinical settings, such as the choice of mammography machine.
The team trained Mirai on the same data set of over 200,000 Massachusetts General Hospital (MGH) exams from their previous work and validated it on MGH test sets, the Karolinska Institute in Sweden and the Chang Gung Memorial Hospital in Taiwan. . Mirai is now installed at MGH, and the team’s collaborators are actively working to integrate the model into care.
Mirai was significantly more accurate than previous methods in predicting cancer risk and identifying high-risk groups in all three data sets. When they compared high-risk cohorts on 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 similarly 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.
“Improved breast cancer risk models allow for targeted screening strategies that achieve earlier detection and less screening damage than existing guidelines,” says Adam Yala, a CSAIL PhD student and lead author of a published paper on Mirai this week Science of Translational Medicine. Our goal is to make this progress part of the standard of care. We work with physicians from 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 across populations and study how best to implement it. .
How it works
Despite the widespread adoption of breast cancer screening, researchers say the practice is controversial: more aggressive screening strategies aim to maximize the benefits of early detection, while less frequent screenings 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 additional imaging and MRI. Some guidelines use age-only risk models to determine if and how often a woman should be screened; others combine several factors related to age, hormones, genetics and breast density to determine further testing. Despite decades of efforts, the accuracy of the risk models used in clinical practice remains modest.
Recently, mammography-based deep learning risk models have shown promising performance. To bring this technology to the clinic, the team identified three innovations it considers critical to risk modeling: joint time modeling, the optional use of non-imaging risk factors, and methods to ensure consistent performance across clinical settings.
1 time
Inherent in risk modeling is learning from patients with different amounts of follow-up and risk assessment at different times: this can determine how often they are screened, whether they should have additional images or even consider preventive treatments.
Although it is possible to train separate models to assess risk at any given time, this approach can lead to meaningless risk assessments – such as predicting that a patient has a higher risk of developing cancer within two years. than within five years. . To address this, the team designed its model to simultaneously predict risks at all times, using a tool called an “additive-hazard layer”.
The additive-hazard layer works as follows: Their network predicts a patient’s risk at one point, such as five years, as an extension of his risk at a previous time, such as four years. By doing so, their model can learn from variable tracking data and then produce self-consistent risk assessments.
2. Risk factors without image
While this method focuses primarily on mammograms, the team also wanted to use non-imaging risk factors, such as age and hormonal factors, if they were available – but did not request them at the time of the test. One approach would be to add these factors as an input element 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 requesting them, the network predicts this information during training and, if it is not there, it can use its own predictive version. Mammograms are rich sources of health information and so many traditional risk factors, such as age and menopause, can be easily predicted from their imaging. As a result of this design, the same model could be used by any clinic globally and, if they have that additional information, they can use it.
3. Constant performance in clinical settings
To incorporate deep learning risk models into clinical guidelines, the models must consistently operate in a variety of clinical settings, and its predictions cannot be affected by minor variations, such as the machine on which 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 various environmental-specific cues. To unlock the model, the team used a contradictory scheme in which the model specifically learns mammographic representations that are invariant to the source clinical environment, to produce consistent predictions.
To further test these updates in various clinical settings, the scientists evaluated Mirai on new test sets from Karolinska in Sweden and Chang Gung Memorial Hospital in Taiwan and found that it performed consistently. The team also analyzed the model’s performance in races, ages, and breast density categories in the MGH test set and in the cancer subtypes in the Karolinska data set, and found that it worked similarly in all subgroups.
“African-American women continue to have breast cancer at a younger age and often in later stages,” says Salewai Oseni, a breast surgeon at Massachusetts General Hospital who was not involved in the work. This, combined with a higher triple-negative breast cancer in this group, led to increased breast cancer mortality. This study demonstrates the development of a risk model whose prediction has remarkable accuracy throughout the breed. The opportunity for clinical use is great. ”
Here’s how Mirai works:
1. The mammogram image is placed through something called an “image encoder.”
2. Each representation of the image, as well as from which view it came, is aggregated with 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 hazard layer predicts a patient’s risk for each year over the next five years.
Mirai’s improvement
Although the current model does not analyze any of the previous results of the patient’s image, changes in imaging over time contain a wealth of information. In the future, the team aims to create methods that can effectively use the patient’s imaging history.
Similarly, the team finds that the model could be further improved by using “tomosynthesis”, an X-ray technique for screening asymptomatic cancer patients. Beyond improving accuracy, further research is needed to determine how image-based risk models can be adapted to different mammography devices with limited data.
“We know that MRI can catch cancer earlier than mammography and that earlier detection improves patient outcomes,” says Yala. But for low-risk cancer patients, the risk of false-positives may outweigh the benefits. With improved risk models, we can design more nuanced risk-screening guidelines that provide more sensitive screening, such as MRI, to patients who will develop cancer to achieve better results, while reducing unnecessary screening and over-screening. treatment for the rest. ”
“We are both 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 radiology and computer science at Emory University, who was not involved in the work. “We study this question extensively and how to detect failure.”
Yala co-authored Mirai with 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 Royal KTH Institute of Technology, Professor Yung-Liang Wan of Chang Gung University, Leslie Lamb of MGH, Kevin Hughes of MGH, lead author and Harvard Medical School professor Constance Lehman of MGH, and lead author and MIT professor 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.