According to the CDC, there are more than 230,000 people diagnosed with breast cancer in the U.S. each year. For these patients, determination of whether cancer cells have metastasized to other parts of the body significantly influences decisions about treatment.
While a pathologist’s report is generally considered the gold standard in the diagnosis of cancer, reviewing diagnostic slides is an extremely complicated task, even for specialists with years of training and experience. Different pathologists can arrive at variable diagnoses for the same patient, which can result in misdiagnoses. Diagnostic agreement for some forms of breast cancer can be as low as 48%. That number is unsurprising, considering the massive volume of information that must be reviewed in order to make an accurate diagnosis.
Dr. Martin Stumpe, Technical Lead, and Dr. Lily Peng, Product Manager at Google Research, wrote in a recent blog post that “There can be many slides per patient, each of which is 10+ gigapixels when digitized at 40X magnification. Imagine having to go through a thousand 10 megapixel (MP) photos, and having to be responsible for every pixel.” Reviewing that much information per slide is time-consuming. Stumpe and Peng worked with a team of Google researchers to address two of the critical issues in the battle against breast cancer: diagnostic variability and pathologists’ time constraints.
The Google research team used pathology images to train machine learning algorithms which have been optimized to detect metastasized breast cancer cells. Imaging analytics driven by machine learning algorithms can identify metastasized breast cancer with higher sensitivity rates than other automated methods and can even rival detection by human pathologists, according to a recent Google research paper.
A HealthITAnalytics article reports that Stumpe and Peng worked with a team of researchers from Google Inc., Google Brain and Alphabet’s Verily Life Sciences (formerly Google Life Sciences) “to apply convolutional neural network (CNN) architecture to a set of training and validation images, surveying the data on a pixel-by-pixel basis. The algorithm produced “heat maps” that predicted the likelihood of tumor cells in a given sample.”
Stumpe and Peng wrote that after customizing “off-the-shelf” deep learning approaches, “including training networks to examine the image at different magnifications (much like what a pathologist does),” Google successfully demonstrated that “it was possible to train a model that either matched or exceeded the performance of a pathologist who had unlimited time to examine the slides.” The improved heatmaps produced by the algorithm reached 89 percent accuracy, compared to 73 percent accuracy from a pathologist with no time constraints who scrutinized 130 slides for 30 hours. Diagnostic accuracy was measured using a localization score (FROC).
Additionally, Stumpe and Peng noted that “These algorithms perform well for the tasks for which they are trained, but lack the breadth of knowledge and experience of human pathologists — for example, being able to detect other abnormalities that the model has not been explicitly trained to classify (e.g. inflammatory process, autoimmune disease, or other types of cancer).” The algorithms should complement the pathologist’s workflow to ensure the best clinical outcome for patients, they explained.
The researchers envision that machine learning algorithms will increase pathologists’ efficiency and diagnostic consistency. They offered examples such as the ability for clinicians to “reduce their false negative rates (percentage of undetected tumors) by reviewing the top ranked predicted tumor regions including up to 8 false positive regions per slide,” and to “easily and accurately measure tumor size, a factor that is associated with prognosis.”