By Gene Munster
Google recently announced a milestone for its applications of deep-learning in medicine. Their algorithm for detecting metastatic breast cancer, Lymph Node Assistant (LYNA), is not only able to positively classify metastatic cancer with 99% accuracy but is also able to accurately highlight the location of the cancerous cells and other suspicious regions within Lymph nodes. In brief:
- Google announced it has developed a deep-learning algorithm capable of more accurately identifying metastatic cancer with 99% accuracy. One recent study found that humans miss metastatic cancer 62% of the time when under time constraints.
- Google’s technology won’t replace Pathologists, who add the critical human element of considering diagnoses in context on a case by case basis, but it offers Pathologists a new and important tool.
- This technology will be to enable doctors to more accurately diagnose conditions in less time.
- We consider this human-in-the-loop automation — a collaboration between human and machine that will make us more productive.
Even though the algorithm by itself is very successful, the best results are shown when pathologists are working with LYNA, suggesting the future is using machine learning to enhance the capabilities of human doctors. Human-in-the-loop automation involves human labor augmented – not replaced – by machines. In this case, a machine would make a first pass. screening samples and flagging possible positives for human review. In a study Google conducted, six board-certified pathologists worked alongside LYNA. The pathologists reported that LYNA made diagnosis easier and, on average, cut the review time per slide in half. This is significant because studies have shown that under time pressure, sensitivity for detecting small metastasis (True Positive / (True Positive + False Negative)) can be as low as 38%. The study also showed a reduction in the rate of missed micrometastases classification by a factor of two.
Limitations Exist, but Results Are Promising
There is still a long way to commercial use and clinical trials will need to be done. Some limitations of the two studies are limited data set sizes and simplified workflow of diagnosis. The studies only looked at one slide per patient but typically multiple will be used. With those limitations in mind, these early results are still very impressive. We are in the early days of integrating deep-learning technology into medicine and are already able to see the dramatic improvement it offers in accuracy. The cost of this accuracy an increase in false positives. This is another reason why pathologists should still be the decision makers of the diagnosis because they have a contextual understanding of these implications that algorithms cannot consider.
Doctors still have contextual awareness and ability to view diagnosis and symptoms on a case-by-case basis to determine the best treatment plan. With the use of machine learning, doctors will be able to navigate this process more quickly and with better accuracy. The most important takeaway from this announcement is that we are getting closer to working versions of deep-learning applications in medicine and that it has a demonstrated power to enhance the ability of doctors.
Disclaimer: We actively write about the themes in which we invest or may invest: virtual reality, augmented reality, artificial intelligence, and robotics. From time to time, we may write about companies that are in our portfolio. As managers of the portfolio, we may earn carried interest, management fees or other compensation from such portfolio.