We think that machine learning is poised to transform medicine, delivering new, assistive technologies that will empower doctors to better serve their patients. Artificial intelligence has dozens of possible application areas, but healthcare stands out as a remarkable opportunity to benefit humanity — and working closely with clinicians and medical providers, Google is developing tools that we hope will dramatically improve the availability and accuracy of medical services.

Medical imaging is one of the areas we are currently exploring in healthcare. Deep learning has already revolutionized the field of computer vision, making practical, in-your-pocket technologies out of what seemed like science fiction just a few years ago. If these new computer vision systems can reach human-level accuracy in identifying dog breeds or cars, might those same systems be capable of learning to identify disease in medical images? Over the last few years, we’ve been working with doctors and clinicians to show that this is indeed possible — not in some far off future, but today. Two of the areas of medical imaging where we’ve made the most progress in our research to date are ophthalmology and digital pathology.

In the area of ophthalmology, we began exploring computer aided diagnostic screening for a disease of the eye called diabetic retinopathy. Diabetic retinopathy is the fastest growing cause of preventable blindness globally. The condition is normally diagnosed by a highly trained doctor examining a retinal scan of the eye, and if caught early, effective treatments are available. If undetected, however, the disease progresses into irreversible blindness. In much of the world, there simply are not enough doctors available to support the volume of screening required to protect the population.

In close collaboration with doctors and international healthcare systems, Google has developed a state-of-the-art computer vision system for reading retinal fundus images for diabetic retinopathy. Our early results are very encouraging, and our algorithm’s performance is on par with U.S. Board-Certified ophthalmologists. We’ve recently published some of our research in the Journal of the American Medical Association and summarized the highlights in a blog post. There is still much work to do to bring the benefits of this research to patients, but ultimately, we hope to help real doctors and clinics expand global screening capacity to cover all at-risk individuals in the world.

In the field of digital pathology, we’ve focused our initial research on algorithms that might assist pathologists in detecting breast cancer in lymph node biopsies. Reviewing pathology slides is a very complex task that requires years of training, expertise, and experience. Even with this extensive training, there can be substantial variability in the diagnoses given by different pathologists for the same patient, which is not surprising given the massive amount of information that must be reviewed in order to make an accurate diagnosis.

To address these issues of limited time and diagnostic variability, we built an automated detection algorithm that can naturally complement pathologists’ workflow and shared our results in this paper and this blog post. In short, we were able to build an algorithm with high sensitivity (92% sensitivity at 8 false positives per slide) that would be complementary to the low false positive rates of pathologists (73% sensitivity at zero false positives per slide).

A closeup of a lymph node biopsy. The tissue contains a breast cancer metastasis as well as macrophages, which look similar to tumor but are benign normal tissue. Our algorithm successfully identifies the tumor region (bright green) and is not confused by the macrophages.

We think these results are just the beginning. There are countless opportunities for machine intelligence to improve the accuracy and availability of healthcare, and we hope that our research will serve as one of many demonstrations of that potential. We’re excited to share future work on healthcare projects in collaboration with doctors and medical systems.

Note: We’re always open to research collaborations on large medical datasets of all types (not just imaging). If you know of an organization that would like to work with us, please fill out this this form.

If you are interested in job or internship opportunities, please visit the Work at Google page.