AI System Outperforms Humans at Diagnosing Lung Cancer

AI System Outperforms Humans at Diagnosing Lung Cancer

Scientists at Feinberg School of Medicine and Google Health Research have developed a deep learning algorithm that outperforms humans at diagnosing lung cancer – The leading cause of cancer-related deaths in the United States.

Each year, lung cancer claims 160,000 lives in the United States. In the latter stages of the disease the cancer is very difficult to treat. Early diagnosis is essential to ensure treatment is provided when it is most effective.

A diagnosis of lung cancer is typically made via X-ray or low-dose CT scan (LDCT), with the latter arguably the best method of detecting the disease. However, while LDCT scans allow lung cancer to be identified, interpreting scans can be problematic. There are many false positives and false negatives, with the former leading to a costly and unnecessary biopsy and the latter delaying the diagnosis. That delayed diagnosis could result in the cancer progressing to a stage when it is no longer treatable.

AI-based systems have been developed which can assist clinicians with the diagnosis of many medical conditions. Now a machine learning algorithm has been developed which can identify lung tumors. Further, the system has been shown to outperform humans.

The AI system was trained using 42,290 LDCT scans taken from the Northwestern Electronic Data Warehouse and other Northwestern Medicine hospital data sources. Training involved the use of the primary LDCT scan and, if available, earlier LDCT scans. By analyzing the earlier scans, it is easier to identify abnormal growth of lung nodules, which helps to determine malignant growths.

The AI algorithm was able to diagnose malignant nodules without human intervention with a high degree of accuracy. To check how the AI algorithm compared to trained clinicians, the researchers presented LDCT scans to 6 board-certified radiologists who had up to 20 years’ experience.

When prior images were available, there was little difference in the performance of the trained radiologists and the trained AI system. However, in the absence of the prior images, the AI system performed better with a reduction of 11% of false positives and 5% of false negatives.

“AI in 3D can be much more sensitive in its ability to detect early lung cancer than the human eye looking at 2D images. This is technically ‘4D’ because it is not only looking at one CT scan but two over time,” explained Dr. Mozziyar Etemadi, assistant professor of anesthesiology at Northwestern University Feinberg School of Medicine in Chicago and co-author of the study.

Not only could the AI system ensure that lung cancer is accurately diagnosed to ensure that patients can receive the treatment they need, it could also ensure that patients avoid costly and risky lung biopsies that they do not need. The AI system could also help to widen screening for lung cancer. At present the vast majority of patients with lung cancer have not yet been screened.

Before the AI system can be used in a clinical setting, it is first necessary to use the system on a much larger sample of data to validate the results.

The study is detailed in the paper – End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography – which was recently published in the journal Nature Medicine. DOI: 10.1038/s41591-019-0447-x

Leave a Reply

Your email address will not be published. Required fields are marked *