The AI algorithm could optimize lung cancer screenings.
Photo/Thumb have been modified. Courtesy of University of Michigan.
A deep-learning model outperformed six U.S. board-certified radiologists in predicting lung cancer, according to research conducted by Google.
The findings of the study, published in the journal Nature, suggest that the algorithm, which uses a patient’s current and prior computed tomography volume, can efficiently predict the risk of lung cancer.
“This work demonstrates the potential for AI to increase both accuracy and consistency, which could help accelerate adoption of lung cancer screening worldwide,” wrote Shravya Setty, a technical lead who works on medical imaging analysis at Google.
The artificial intelligence (AI) model performed at a 94.4% area under the curve on more than 6,700 National Lung Cancer Screening Trial cases. The model performed similarly on an independent clinical validation set of more than 1,100 cases.
When the algorithm did not have prior computed tomography imaging available, the model outperformed six radiologists with absolute reductions of 11% in false positives and 5% in false negatives.
The model performed similarly to radiologists when it used a single CT scan for diagnosis.
While lung cancer screenings are valuable, Shetty wrote that only 2 to 4% of eligible patients in the U.S. are screened.
Diego Ardila, senior software engineer at Google AI, and his research team used three key components in their approach.
First, the team constructed a 3D convolutional neural network (CNN) model that performed end-to-end analysis of whole-CT volumes using low-dose computed tomography volumes with pathology-confirmed cancer as training data.
Next, the researchers trained a CNN region-of-interest detection model to find 3D cancer candidate regions in the CT volume.
Then, they developed a CNN cancer risk prediction model that operates on outputs from the cancer region-of-interest detection model and full-volume model. The team also trained the CNN cancer risk prediction model on case-level, pathology-confirmed cancer labels.
The team leveraged more than 48,800 de-identified chest CT screening cases from the National Institute of Health’s research data set from the National Lung Cancer Screening Trial study and Northwestern University.
The research team validated the results with another data set and compared the findings against six U.S. board-certified radiologists.
“We created a model that can not only generate the overall lung cancer malignancy prediction (viewed in 3D volume) but also identify subtle malignant tissue in the lungs (lung nodules),” Shetty wrote.
The software engineer added that while the results are encouraging, further studies will assess the model’s impact and utility in clinical practice.
The research team is working with Google’s cloud healthcare and life sciences teams to serve the model through the Cloud Healthcare API. Investigators are speaking with partners to continue clinical validation research and deployment.
Get the best insights in digital health directly to your inbox.
Deep Learning Model Outperforms Clinical Model in Lung Cancer Survival Prediction
AI Diagnoses PTSD with High Accuracy