Lung cancer staging can be challenging for humans, who are known to have difficulty distinguishing adenocarcinomas from squamous cell carcinomas, for example.
In the study, a group from Stanford University Medical Center examined more than 2,000 images from the Cancer Genome Atlas, which included data on the grade and stage of either adenocarcinoma or squamous cell carcinoma. The researchers used the images to train a computer software program to identify nearly 10,000 cancer-specific characteristics, whereas experienced human pathologists normally can distinguish a few hundred traits.
The researchers focused on a subset of cellular characteristics identified by the software that could differentiate tumor cells from the surrounding noncancerous tissue, identify the cancer subtype, and predict survival after diagnosis for each patient. The machine-selected characteristics included not just cell size and shape, but also the shape and texture of their nuclei and the spatial relations among neighboring tumor cells, the authors wrote (Nat Commun, August 16, 2016).
The computers assessed even tiny differences across the thousands of samples many times more accurately and rapidly than a human, the study concluded. The results were validated based on the software's ability to distinguish short-term survivors from those who lived significantly longer in a separate dataset of 294 lung cancer patients from the Stanford Tissue Microarray Database. In addition to predicting patient survival times better than the standard approach, the machine-learning approach accurately differentiated between the two types of lung cancer.
Pathology as it is now practiced is highly subjective, commented Michael Snyder, PhD, professor and chair of genetics at Stanford University, in a statement. Two skilled pathologists evaluating the same slide will only agree about 60% of the time, he said. This approach replaces subjectivity with quantitative measurements that are likely to improve patient outcomes.
Copyright © 2016 AuntMinnie.com