AI Could Help Radiologists Make Breast Cancer Diagnoses
IBM Research developed an AI model to study mamograms and offer a "second reader" for radiologists.
By Jim Utsler01/02/2020
If properly curated, AI has the potential to be applied to many different use cases. For example, IBM Research-Haifa recently announced the findings of a research study about an AI model that can predict which patients are most likely to develop breast cancer within a year—at rates acceptable for breast-cancer screening by human radiologists.
And this is no trivial matter. According to the American Cancer Society (bit.ly/2D3fY6t), breast cancer is the most commonly diagnosed cancer among women—and the leading cause of cancer-related deaths in women—across the globe.
A Deeper Look
For years, digital mammography has been the main imaging method for breast-cancer screening, with recommendations typically ranging from one to two years, depending on age and personal and familial history. The results of the exam are then interpreted by radiologists who scrutinize the images for the existence of malignant findings.
But as Michal Chorev, research staff member, IBM Research, notes in a recent blog post, “Analyzing mammograms is a challenging task. The differences between lesions and background could be very subtle: There are multiple types of possible findings which differ from each other in shape, size, color, texture and other factors. A second reading of mammograms by an additional radiologist has been proven to increase sensitivity and specificity. However, a lack of trained radiologists and time limitations often makes it difficult to incorporate second readers as part of the standard screening procedure in many countries.”
To help address this issue and reduce breast cancer mortality rates, an IBM research team in Haifa hypothesized that an AI model could be applied to assess breast cancer at a level acceptable for breast cancer screening radiologists. In cooperation with Maccabi Health Services, a health provider with 2.3 million members, and Assuta Medical Centers, the largest chain of hospitals in Israel, they were able to work with clinically collected and anonymized mammography images that were also linked to holistic clinical data and biomarkers for each patient.
“This astonishing amount of data provided a deep pool of information from which our machine learning models could learn, and allowed these algorithms to connect patterns and trends that may not have been possible otherwise,” Chorev notes.
Training the System
Using this information, the team created a new algorithm that incorporates mammograms and comprehensive electronic health records for use in the prediction of potential breast cancer. Built on deep learning models, they trained this system to achieve, as defined by the Breast Cancer Surveillance Consortium benchmark for screening digital mammography, an accuracy comparable to radiologists.
In a retrospective study, the team collected a data set of nearly 53,000 images from more than 13,234 women who underwent at least one mammogram between 2013 and 2017 and also had health records for at least one year prior to the mammogram.
According to Chorev, “We trained our algorithm on 9,611 mammograms and health records of women, with two objectives: to predict biopsy malignancy and differentiate normal from abnormal screening examinations. We created a combined machine- and deep-learning model for those objectives, using as input the mammography’s standard four-view images and the detailed clinical histories.”
Using both imaging data and comprehensive patient health histories, the model correctly interpreted 77% of noncancerous cases while predicting the development of breast cancer in 87% of analyzed cases.
“More research needs to be conducted, but it’s possible that more accurate prediction could hold the potential to reduce the number of women being sent for unnecessary tests—or experiencing the trauma of being needlessly assigned as high risk—by traditional models,” Chorev remarks. “Combining clinical data with imaging information could offer the potential to more accurately validate initial results.”
For more information
The results of this research were published in the paper “Predicting Breast Cancer by Applying Deep Learning to Linked Health Records and Mammograms,” which appeared in the scientific journal Radiology in June 2019.
Jim Utsler, IBM Systems magazine senior writer, has been writing for IBM since the mid-1990s.