Google: Deep learning algorithm for cancer detection more effective than pathologist

Google has developed an algorithm to support pathologists in diagnosing breast cancer. Through deep learning analysis of images, this method is as effective as or more effective than a pathologist.

Google presented the results of its research Friday, along with a paper. The system aims to detect tumors of 100×100 pixels on 400 images of 100,000×100,000 pixels. The images used of human lymph nodes at forty times magnification come from the Radboud UMC in Nijmegen. This made it possible to identify 89 percent of the tumors present, the researchers say in the paper. A human pathologist scores lower, with a rate of 73.2 percent if no time limit is set.

The algorithm was trained with small parts of the images, after which a heat map was made of the probability of the presence of cancer cells. The system also worked with images from other hospitals, which had been scanned with different equipment. However, it is unable to perform tasks for which it has not been trained, such as detecting other abnormal phenomena.

On the left the scanned image with on the right the tumor in white and generated heatmaps

According to Google, the research should help support the work of pathologists by increasing their efficiency and consistency. For example, a pathologist could reduce his percentage of undetected tumors by looking at the system’s predictions. Jeroen van der Laak, researcher at Radboud UMC, tells the NOS about the research: “Analyzing lymph nodes is time-consuming and repetitive work. You have to work very concentrated and must not miss anything. A computer provides constant quality and can store large amounts of data. process quickly.”

In Nijmegen, pathologists already work with their own algorithm that differs from the Google variant. Van der Laak further says that it does not have the same precision, but that improvements may be possible based on the research presented on Friday.