Stanford researchers recently developed a computer model capable of both analyzing microscopic breast cancer images and offering patients a prognosis with unprecedented accuracy and consistency.
The Computational Pathologist system (C-Path) is able to classify the types of cancer cells present in a patient and the level of aggression of the cells. The system also identifies key features in tumor tissue that may indicate chances of survival. The model’s computer analyses were a significant improvement in statistical accuracy over those carried out by human pathologists using the same data.
Stanford researchers trained the computer model using tumor tissue images from 248 breast cancer patients whose survival data for the subsequent five years were known to the researchers. C-Path examined 6,642 features in the cancer image before independently identifying the most critical factors in determining survivability. The researchers then created a scoring system to predict patient outcome.
C-Path also identified structural features in the tumor tissue images as more important than previously thought by pathologists in determining patient outcome.
“We built a model based on features of the stroma — the microenvironment between cancer cells — that was a stronger predictor of outcome than one built exclusively from features of [cancerous] epithelial cells,” said Andrew Beck, lead author of the study. “The stromal model was as predictive as the model built from both stromal and epithelial features.”
Daphne Koller, a professor in the School of Engineering and the study’s senior author, said in an email to The Daily that the study’s results were “surprising and significant because today’s cancer grading scheme looks only at the cancer cells.”
“This finding supports the emerging view that cancer isn’t just a bunch of cells gone awry, but rather an entire ecosystem,” she said.
Beck, now an assistant professor of pathology at Harvard Medical School, originally undertook the study as part of his post-doctoral work at Stanford. His research focused on the concept of prognosis through algorithm, which stemmed from collaboration with Koller. Beck’s experience in pathology, combined with Koller’s expertise in machine learning and image analysis, allowed for the study to develop over a two-year period.
The study confirmed its results using a “validation” set of data from 328 women. The nature of the subject material also complicated efforts to build the model. Among the challenges researchers faced was converting images to quantitative data.
Beck noted that there remain several obstacles to deploying the computer model for clinical prognoses, but anticipated broader use of the model within a few years. He indicated the model should be tested with larger images than the microscopic tissue slides used in the study. Beck added that the algorithm has to be able to adapt to the different tissue staining procedures used by different medical institutions.
Another challenge remaining is the absence of lab standardization for the C-Path system. At the moment, researchers have to retrain the model’s epithelial and stromal tissue indicators when changing between data sets from institutions with different staining protocols.
In the future, the researchers anticipate the model will be applied not only to other forms of cancer but also to other diseases. The self-teaching algorithm that C-Path employs could train itself to respond to different training data and parameters and could thus formulate different predictions accordingly.
“One particularly exciting application is to train the model on a cohort of patients treated with a particular diseas . . . and predict which patients will respond to the drug and which won’t,” Koller said. “This would allow the system to be used directly in helping guide clinical care.”
By improving the accuracy and consistency of breast cancer prognoses, the model may allow doctors to make more informed decisions and to tailor treatments to the severity of the patient’s condition. While the model seemingly offers the possibility of advanced prognoses in areas traditionally devoid of expert medical coverage, researchers intend the system to act as a complement for pathologists.
“This [model] will never replace a pathologist, but will eventually emerge as a decision support tool,” Beck said. “It will be one that delivers improved information and results for doctors and patients everywhere.”