It has been said pathology is subjective, and it is well known that pathologists can differ wildly when analyzing the same tissue samples. Now, it would seem that based on a new study that computers can assess slides of lung cancer tissue with greater accuracy than scientists. This leads to greater accuracy in classifying tumors and predicting individual prognoses.
Overall the computers could differentiate with greater accuracy not only between two types of lung cancer but also have greater accuracy in predicting survival times and tumor classification.
This study assessed the viability of the approach in terms of lung cancer, but scientists believe that the principles could be applied to other forms of cancer as well. Computers are able to give a deeper understanding of where the molecular mechanisms of cancer are concerned with the programmable ability to connect cancer’s pathological “features” with possible outcomes based on established data.
Additionally, pathologists have employed a light microscope to determine the grade of particular cancer by examining thinly sliced cross-sections of tumorous tissue. The greater the amount of “abnormal” tumor tissue that was recorded, the higher the grade of cancer.
This “grading” of cancer is essential as it is utilized in predicting how an individual will be affected by cancer as well as being a determining factor regarding exactly to treat cancer. Alas, this system is not consistently accurate when assessing lung cancer. For example, subtypes that occur in lung cancer, such as adenocarcinoma and squamous cell carcinoma, are notoriously hard to differentiate between. There is also the problem that the grade and stage of lung cancer patient’s cancer is not always an accurate indicator of prognosis.
In all, scientists used over 2000 different images from The Cancer Genome Atlas (CGA). These were images of individuals who had adenocarcinoma or squamous cell carcinoma. The CGA information contained the grade and stage of each specimen and survival rate of the patient associated with each specimen.
Scientists then took these images and taught the software to correctly identify 10,000 specific aspects of cancer represented in the images that were beyond to capability of the human eye to recognize. Taken into account were nuclei texture-shape and the space between tumor cells in the same proximity. In comparison, the average pathologist uses a few hundred aspects-characteristics to generate their grade and stage outcomes.
Scientists then focused on a sub-category of cell characteristics the software had isolated – being able to tell the difference between tumorous and noncancerous tissue within close proximity to one another, identification of cancer sub-types, and predictive ability in terms of individual survival.
Being able to identify these hitherto unidentified characteristics could vastly improve the prediction of severity and survival rates. It can also give a deeper understanding of the processes that trigger cancer at a molecular level.