A Dartmouth research team is harnessing machine learning technology to predict malignant breast cancer lesions. Saeed Hassanpour, assistant professor of biomedical data science and epidemology at the Geisel School of Medicine, and his team are focused on developing this technology to predict the possibility that a breast lesion found during medical examinations is or will become cancerous.
Hassanpour said that breast cancer screenings are widely used, but can induce a false positive, which put women in danger of overdiagnosis and overtreatment.
He explained that typically, if a lesion is found after a mammography, doctors perform a core needle biopsy on the patient. If a marker for high risk breast cancer incidences, known as atypical ductal hyperplasia, is found, surgery is performed to determine whether the lesion is malignant or benign, according to Hassanpour.
“Seventy to 80 percent of women didn’t need this surgery,” Hassanpour explained. “Only 20 to 30 percent of patients are [found to have cancerous lesions].”
Following this discovery of overdiagnosis and treatment, Hassanpour and his colleagues searched for less-invasive alternatives for women seeking diagnoses.
“We thought it would be good to introduce a personalized decision-making approach” he added.
Hassanpour said that he and his team want to help women who do not need surgery avoid the distress, costs and potential side-effects of an unnecessary operation.
“If patients demonstrate they have a lower risk of cancer, they can decide if they want to go through with the surgery or opt for an alternative, less invasive surveillance” Hassanpour said.
Hassanpour said his lab consists of students from the quantitative and biomedical sciences PhD and Masters program and the computer science PhD, Masters and undergraduate programs. Other members include research scientists and post-doctoral fellows with backgrounds in computer and data science.
“My lab works in developing new machine learning and natural language processing methods to distill health-related insights from bio-medical data” Hassanpour explained. “This is mostly unstructured data like medical records or images.”
Roberta DiFlorio Alexander, associate professor of radiology and gynecology and obstetrics at the Geisel School of Medicine, was approached by Hassanpour to collaborate with the research team. She said she was very excited to learn how AI might be applied to breast imaging.
Hassanpour said he was aware of DiFlorio Alexander’s research in this field and reached out to her to see if she was interested in working together.
In an email statement, DiFlorio Alexander wrote “[Hassanpour’s] extensive knowledge in this field allowed him to understand that we now have the computational ability to analyze big data and combine computer-extracted information with human-extracted information to improve our understanding and evaluation of breast cancer and other medical conditions.”
DiFlorio Alexander highlighted the use of artificial intelligence in radiology research. She noted that the use of digital imagining allows researchers to mine vast amounts of data, that when combined with molecular and proteomic data could facilitate personalized medicine and a better understanding of individual disease processes.
“This would open up the possibility of targeted, personalized treatment for each patient’s unique cancer,” DiFlorio Alexander wrote.
Hassanpour’s project is an off-shoot of research started at the College by computer science professor Lorenzo Toressani.
Arief Suriawinata, professor of pathology and laboratory medicine at Geisel, said he approached Toressani and the two began collaborating. Suriawinata said soon after Hassanpour began working at the College and he joined their research team. Hassanpour’s expertise and his team of undergraduate and post-graduate researchers helped accelerate the project in the past few years, according to Suriawinata.
“Colo-rectal screening is universal these days,” Suriawinata explained.
Often, clinicians identify pre-malignant colo-rectal polyps that, if not removed, are likely to become cancerous in a few years, according to Suriawinata. The identification of these colorectal polyps is similar to the identification of ADH in Hassanpour’s study.
“Studies have shown each pathologist may have some variability in making the diagnosis” Suriawinata said.
Instead of using slides alone to identify colorectal polyps, Suriawinata said that machine learning and a standard algorithm could make diagnoses more consistent in the future.
“Roberta and I are working on exchanging this model and developing machine learning models which can include genetic bio-markers [and clinical factors],” Hassanpour said, adding that their team collected data from medical records of all the patients identified as positive for ADH by a core needle biopsy at DHMC, dating back to 2011.
The team wants to build comprehensive risk assessment for breast cancer patients and eventually extend the machine learning model for the diagnosis, prognosis and treatment of other types of cancer.
“We want to extend this work in multiple directions,” Hassanpour said. “We want to expand the work to include other [noncancerous but] high risk breast lesions … and hopefully deploying the model in the clinical practice.”
