A new algorithm could significantly eliminate unnecessary operations for breast cancer lesions by predicting which ones will become cancerous.
Researchers from Massachusetts General Hospital (MGH), Harvard Medical School, and Massachusetts Institute of Technology (MIT) have developed a new machine learning model that could predict which high-risk lesions will advance to cancer and need surgical intervention and which lesions are low-risk. Identifying the lesions that will not require surgery could help clinicians reduce unnecessary operations, according to the authors. Their study was published in Radiology.
“Currently, the decision about whether or not to proceed to surgical excision is usually based on the histologic subtype of the high-risk lesion. However, treatment of patients based on histologic subtype alone has led to variable and sometimes aggressive treatment,” said Manisha Bahl, MD, MPH, lead author of the study, radiologist at MGH, assistant professor of radiology at Harvard Medical School, and director of the MGH Breast Imaging Fellowship Program, told MDAlert.com. “We used machine learning to develop a more robust risk-stratification approach that could be used to guide clinical decision making.”
.jpg)
MRI scans. (Source: jber.jb.mil)
According to Dr. Bahl, machine learning refers to algorithms that can be designed to evaluate and make predictions on the basis of new and complex features. She said that radiologists have taken an interest in machine learning because it has the potential to improve clinical practice by detecting hemorrhages in the brain, differentiating between cancerous and noncancerous lung nodules, or segmenting tumors in the liver.
In regards to high-risk breast lesions, Dr. Bahl said that the algorithm can help clinicians and patients navigate treatment based on the severity of the lesion. “Machine learning allows us to incorporate the full spectrum of diverse and complex data that we have available, such as patient risk factors and imaging features, in order to predict which high-risk lesions are likely to be upgraded to cancer and, ultimately, to help our patients makes more informed decisions about surgery versus surveillance,” she said.
Dr. Bahl and her colleagues first identified women who underwent image-guided core needle biopsy from 2006 to 2015 at MGH and had been diagnosed with a high-risk lesion. They then developed a structured database that included patient information, such as risk factors for breast cancer, mammographic findings that led to biopsy and high-risk lesion diagnosis, and results from their biopsy.
Researchers from MIT, whom Dr. Bahl and her colleagues collaborated with, developed a machine learning model using a “random forest classifier”, which used data elements including core biopsy pathologic report text, patient risk factors, and histologic results from 700 high-risk lesions to analyze over 300 high-risk lesions. The model then put out a score that reflected the likelihood that the lesion would advance into cancer at surgery and compared the model’s performance with outcomes from surgical excision versus surveillance of high-risk lesions.
They found that if the high-risk lesions that were categorized as low-risk for cancer upgrade and observed while the rest were surgically excised, 97.4% of malignancies would have been diagnosed at surgery and 30.6% of surgeries of benign lesions could have been avoided.
“We were surprised that our model performed better than the traditional strategy of surgically excising only certain histologic subtypes of high-risk lesions, such as atypical ductal hyperplasia, lobular carcinoma in situ, and atypical lobular hyperplasia,” Dr. Bahl said. “Compared to the traditional strategy, use of our model would increase the number of cancer detected and decrease the number of benign surgeries performed.”
There was one case of cancer upgrade that was misclassified by the machine learning model, Dr. Bahl said. A 34-year-old woman with a history of Cowden syndrome was diagnosed with a high-risk lesion at core needle biopsy that was upgraded to ductal carcinoma in situ (DCIS), or noninvasive cancer, at surgery. The machine learning model was not able to recognize rare genetic syndromes such as Cowden syndrome, so it did not provide a score high enough to let clinicians know that the woman’s lesion was high-risk, not low-risk.
Dr. Bahl and her colleagues are currently working to use the risk prediction tool in patients with DCIS and to incorporate mammographic images and histopathology slides into the machine learning model.
“The application of machine learning to breast imaging is in its infancy, and I’m excited about the tremendous potential of machine learning to improve clinical decision making and decrease the morbidity and costs of overtreatment,” Dr. Bahl said.
