Google AI Detects Lung Cancer Faster than Doctors
On Tuesday, Google announced that it has trained an artificial intelligence to recognize early symptoms of lung cancer much faster than oncologists can. This has the potential to greatly improve the survival rate of lung cancer patients, as early detection of the disease has been shown to be crucial to patients’ survival in past studies. This was revealed on Tuesday at Google’s annual I/O Developer Conference, hosted at the Shoreline Amphitheatre in Mountain View, CA, from May 7-9.
Computed Tomography (CT) is a process which produces a cross-sectional image of an object by combining multiple layers of X-ray measurements.The AI was trained through machine learning to spot subtle lesions (damaged tissue, in this case caused by yet-unseen tumors) in the lungs using sets of CT scans from the National Cancer Institute and Northwestern University.
The AI was found to be effective at a level at least equal, if not superior, to a trained radiologist. It was reported in the presentation that five of six trained radiologists fail to spot these lesions. In one case, a patient was diagnosed with late-stage cancer a year after a supposedly clean CT scan. However, the AI was able to detect signs of the cancer in this initial scan. Dr. Lily Peng, Product Manager at Google’s AI research and the development arm, Google Brain, said that detecting cancer a year before an official diagnosis “could translate into an increased survival rate of 40 percent.”
The simplest method of machine learning involves repeatedly testing algorithms to recognize certain patterns in data; in this case, lung lesions in chest CT scans. This process is random at first, as an algorithm has no inherent method of “knowing” what it looks for, only that it found the pattern correctly or incorrectly. Over time, the algorithms that are most successful in finding the pattern are kept, while the others are discarded; more random alterations are made to the code, and testing begins again. Eventually, enough of these changes can be built up over time so that a network of interacting lines of code forms an algorithm which is capable of spotting the pattern with a high degree of consistency. With all that being said, it is important to note that, no matter how complex a neural network becomes, it is only able to perform the singular task that it has been tested for.
Deep learning -- the specific variant of machine learning used to develop this AI -- relies on multiple layers of pattern detection which interact with one another. The algorithm learns to abstract the data from each layer to form a more complex method of pattern recognition. For example, a series of pixels in an image can be abstracted into an edge, which can be further abstracted into arrangements of edges which form increasingly complicated shapes. This allows for more advanced forms of pattern recognition, some of which are in use by other Google products and services, such as speech recognition and live translation.
If the new AI is as effective as Google claims it to be, this technology carries large implications for the medical field. Cancer is, according to the World Health Organization, the leading cause of death worldwide. Among the various forms of the disease, lung cancer is deadliest, accounting for 1.76 million deaths of out a total of 9.6 million in 2018. The WHO also notes that, if detected early, more effective and possibly even curative treatments can be administered. If diagnosed too late, these treatments are rendered ineffective, and the mortality rate would increase as a result.
Dr. Peng noted in her presentation that over 80% of lung cancers are not caught early. If this AI is able to detect the presence of lung cancer in early CT scans, hundreds of thousands of deaths could be prevented from lung cancer alone. Dr. Peng said that what they’ve seen so far was a “promising but early result,” and that Google was “very much looking forward to partnering with the medical community to use technology like this to improve outcomes for patients.”