This AI can diagnose early lung cancer, and human doctors are afraid of being unemployed?

Today, in a recent paper published in Nature Medicine , scientists from the Google Health Research Department and the Northwestern University School of Medicine have teamed up to bring an artificial intelligence system that can be based on the chest. CT scan, detection and analysis of malignant pulmonary nodules, in order to early diagnosis of lung cancer. It is even more accurate than radiologists.

Lung cancer is currently the world's highest morbidity and mortality cancer, and the number one killer of cancer patients in World. Large-scale clinical trials have shown that low-dose CT scans of the chest can identify important features of lung cancer, reducing mortality by 20% to 43% , which is the preferred method of screening for lung cancer.

However, the accurate detection of nodules is a challenge – the screening method has a high error rate and therefore has limited practicality. Considering the influence of other clinical factors, many lung cancers are advanced at the time of discovery and are difficult to treat effectively. How to improve the accuracy of early diagnosis of lung cancer is a problem that scientists want to make AI (Artificial Intelligence) give full play to its advantages.

To this end, Google's AI scientists have developed a deep learning model. This is an AI technique that teaches computer learning through examples.

The researchers used more than 40,000 CT scans from nearly 15,000 patients to train AI, and nearly 600 of them were confirmed to have cancer by biopsy within one year.

“Radiologists typically check hundreds of two-dimensional cross-sections in a single CT scan, but we let the new machine learning system look at the lungs in a huge three-dimensional image.”

Co-author of the study, Mozziyar of Northwestern University Professor Etemadi introduced.

The growth rate of suspected pulmonary nodules is an important indicator of malignant tumors. In order to allow AI to predict the degree of malignancy of lung nodules without human involvement, the investigator not only prepared a CT scan of the patient when he was first diagnosed. The input was also compared using an earlier CT scan.

"The AI not only looks at the current CT scan, but also compares the previous scan results, so in theory, this is a '4D' scan," added Professor Etemadi.

Image From nature.com

The accuracy of this deep learning model was tested in 6716 test cases. The results showed that it was able to detect very small malignant pulmonary nodules with 94% accuracy.

At the same time, AI performed a "competition" with six radiologists with an average of 8 years of clinical experience. In the case of a previous CT scan image, the AI system is comparable to the radiologist's. And in the absence of a previous CT scan image of the situation, AI and even it surpasses the performance of all six radiology medical experts, 11% reduction in false positive and false negative 5% reduction . High sensitivity and low miss rate means that if used in a clinical setting, the additional burden of unnecessary follow-up can be reduced while the tumor is missed less.

Although the authors caution that this model requires extensive clinical validation, this finding continues to show hope for improved management and prognosis in patients with lung cancer. With AI's strong learning ability, we look forward to clinically proven, such a system can help doctors to assess lung cancer screening as soon as possible, for patients to take lung cancer as soon as possible.