Google has achieved a milestone in machine learning research that will boost the company’s broader ambitions in healthcare. In a paper published today in Nature Medicine, Google researchers present an end-to-end deep learning model that can predict lung cancer comparably or better than human radiologists. Researchers from Northwestern Medicine, Stanford Health Care and Palo Alto Veterans Affairs, and New York University also contributed to the research.
The World Health Organization identifies lung cancer as the most devastating cancer globally, and the major factor contributing to the 1.8 million annual deaths as poor prognosis.
Imaging tests are the most common method for diagnosing lung cancer, wherein doctors and radiologists examine a patient’s chest X-rays or Computed tomography (CT) scans. The American Lung Association advises that early detection by low-dose CT screening can decrease lung cancer mortality by 14 to 20 percent among high-risk populations.
Unfortunately, the US Centers for Disease Control reports that less five percent of patients who met the screening criteria actually got screened. The screening rates can be even lower in areas where healthcare resources are in short supply. Google wants to change that by leveraging AI to improve lung cancer screening and make it more accessible. CEO Sundai Pichai tweeted “today we’re publishing our work in @NatureMedicine showing how these methods could boost chances of survival for many people at risk around the world.”
Google’s lung cancer prediction model was built and trained on TensorFlow and comprises two frameworks: a full CT volume model to generate lung cancer malignancy predictions (viewed in 3D volume), and a malignant lesion detection model to identify subtle malignant tissues in lung nodules. The model can also take previous CT scans as inputs to improve its prediction accuracy.
This study used three datasets: LUNA, LIDC, and NLST. The model was trained on almost 46,000 de-identified chest CT screenings from National Institutes of Health’s (NIH) research dataset.
Researchers conducted experiments with and without using previous CT scans. In the first experiment using a single CT scan for diagnosis, the Google model detected five percent more cancer cases and reduced false-positives by 11 percent compared with a human group of six unassisted radiologists.
The model also achieved a state-of-the-art performance (94.4 percent area under the curve) on 6,716 National Lung Cancer Screening trial cases. In the second experiment, which used previous CT scans, the Google model was comparable to the radiologists’ performance.
Google researchers say the results “show the potential for deep learning models to increase the accuracy, consistency and adoption of lung cancer screening worldwide.”
Annual global healthcare spending will reach US$8.7 trillion dollars by 2020, according to Deloitte. Although Google is a relative newcomer to the healthcare industry, the Silicon Valley tech giant has the resources and the motivation to apply the power of AI in this growing space. Google research arm Google Brain has already developed deep learning algorithms to improve grading of prostate cancer, detect metastatic breast cancer, and predict a patient’s risk of heart attack or stoke.
While Google’s lung cancer prediction study remains in the research stage, its deployment in hospitals could be expected within several years. In 2016 Google introduced its deep learning studies in predicting blindness-causing diabetic retinopathy from patients’ retinal photographs, and this revolutionary AI system is already in use in an Indian eye hospital where it helps doctors with detection and diagnosis.
The paper End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography was published in the Nature Medicine Journal.
Journalist: Tony Peng| Editor: Michael Sarazen