Google Develops Deep Learning-Based Artificial Intelligence for Early Lung Cancer Detection

Low-dose computed tomography screening for lung cancer is known to significantly reduce mortality (death rates) among people at high risk because of heavy smoking. However, low-dose lung computed tomography also has disadvantages: annual exposure to radiation, even at a low dose; overdiagnosis; false positives (noncancerous tumors being mistaken for cancer, leading to biopsies); and operations.

Artificial intelligence includes the subset of deep learning, which may help reduce these risks, according to a study in Nature.

In deep learning systems, human-designed algorithms are responsible for collecting, sorting and analyzing large amounts of data. These large amounts of data (big data) feed an artificial neural network (a set of algorithms modeled after the human brain) and are then interpreted. Programmers do not tell the machine what to do with the data; instead, the machine analyzes the data and makes decisions based on what it has learned. In most cases, the more data it receives, the more accurate its decisions become.

Researchers at Google developed an artificial neural network containing several processing layers. They fed this multilayer artificial neural network lung computed tomography (CT) scan data from people, some of whom had lung cancer, some did not, and some had nodules that developed into cancer.

The technology showed an extraordinary ability to detect lung cancer. When compared with 6 radiologists, the system achieved false-positive (interpreting a finding as cancer when it was not cancer) reductions of 11%, false-negative (interpreting a finding as benign when it was cancer) reductions of 5%, and thus performed very well. When allowed to see previous CT scans, the algorithm produced results equivalent to those of expert radiologists. Overall, this artificial intelligence model was 94.4% accurate.

Although Google's deep learning system is not ready for widespread use, the results offer insight into the potential role of artificial intelligence in medicine. Ultimately, this application creates an opportunity to optimize CT screening for early lung cancer detection through computer assistance and automation. The fact that machines are producing better results in medical tests every day and supporting clinical decisions is a very positive development. However, the low awareness of lung cancer screening among most people who smoke and the continuing challenge of prevention remain separate problems. Automating screening through clinical decision support systems such as Google's deep learning model has the potential to increase the accuracy, consistency and adoption of lung cancer screening.