Machine Learning Tool Predicts Estrogen Receptor Status
Although it’s not ready for prime time, deep neural network might one day predict responses to hormonal therapy for breast cancer.
Estrogen receptor is a protein found inside the cells of female reproductive tissue, some other types of tissue and some cancer cells. If cancer cells have estrogen receptors, they may need estrogen to grow, and this may affect how the cancer is treated. In order to categorize a breast tumor, determine prognosis and select treatment regimens, doctors in the U.S. typically take biopsies of the tumor and stain it to make estrogen receptors and other types of cellular receptors visible under a microscope. But these immunohistochemistry assays may be inconsistent across laboratories, and they are somewhat expensive.
In a study published Sept. 4 in NPJ Breast Cancer, researchers introduced a machine learning framework to predict clinical estrogen receptor status -- defined as greater than 1 percent of cells positive for estrogen receptor by immunohistochemistry staining -- from the spatial arrangement of nuclear features. The learning pipeline segments nuclei from labeled tissue images, extracts their position, shape and orientation, and then passes them to a deep neural network to predict estrogen receptor status.
After training on tissue cores from 57 patients with invasive ductal carcinoma -- the most common type of invasive breast cancer -- the pipeline predicted estrogen receptor status in an independent test set of samples from 56 patients. Specifically, the network learned that estrogen receptor-negative tumors were associated with larger nuclei that exhibited more variability in their features.
The pilot study included a relatively small sample size, and this test is not close to being a replacement for immunohistochemistry at this early stage. But the researchers envision that deep learning could one day be used to predict responses to hormonal therapy.