In short: Google’s latest study into the use of its AI-based Lymph Node Assistant (LYNA) highlights how the AI might be used in the real world, the search giant revealed via its Google AI Blog. The research centered around tasking human pathologists to diagnose metastatic breast cancer in lymph node slides. According to participants in the study, the use of LYNA to detect the cancer cells to be diagnosed made the process much easier. In fact, the company says it cut the review time by around half for each slide used.
Background: The new study, meanwhile, is actually a followup on research into the use of AI for cancer detection that was first reported at the beginning of last year. For the previous study, Google was able to show AI had several advantages over pathologists in terms of detecting cancer. Specifically, the study examined lymph nodes and utilized AI algorithms to examine gigapixel-sized pathology slides, showing images for diagnosis containing billions of pixels. The AI was shown to be able to detect problems down to a scale of only 100 x 100 pixels. In terms of prognosis, the company showed that its AI could reduce false negatives by up to 75 percent. Its overall accuracy at detecting cancer was at around 99-percent.
However, its classification of the stage of detected cancer was not as accurate as had been hoped, Google says. Subsequent studies showed that in as many as a quarter of cases, secondary review by a human pathologist showed the classification to have been incorrect. Furthermore, time constraints impacted detection sensitivity for small metastases. In effect, the detection of cancer itself was accurate but the AI is not able to classify what stage the cancer is at. That determination plays a significant role in determining the viability of treatment options not only in breast cancer but in a wide variety of cancers – forcing the company to re-examine how LYNA might best serve the medical community. What the study really highlights is one of the biggest benefits that AI can bring to humanity, not just in terms of disease detection but more broadly. As noted in Google’s research, the current conclusion is not that AI is best at fully identifying cancer but as an augmentation for pathologists in quickly identifying cancer earlier.
Impact: That’s because LYNA can still accurately identify cancerous growth more easily than a pathologist, who may suffer some loss in accuracy due to fatigue the sheer number of slides viewed. AI would not experience that same limitation, although the ultimate prognosis and treatment plan are still going to vary from patient to patient and that’s something that humans are still needed to determine. That trend toward AI as an assisting technology rather than a replacement for real people is present in the vast majority of industries where AI is being seriously considered – with the possible exception of self-driving cars. In terms of medicine, the use of AI could open opportunities for more widespread and routine cancer screening processes that are more accurate at earlier stages. That would leave pathologists to focus on deeper diagnostics and better care plans.