AI further promotes LDCT to predict lung cancer

Using deep learning algorithms on low-dose chest CT helps radiologists accurately estimate a patient’s risk of whether the identified lung nodule is malignant.

Low-dose CT (LDCT) is effective in screening individuals at high risk of lung cancer, such as long-term smokers, and the number of people undergoing these scans is increasing. However, the correct distinction between cancerous and benign nodules remains an important issue and influences treatment decisions, so accurate evaluation is important.

In an article published on May 18th Radiology, Dutch researchers detail artificial intelligence (AI) tools that outperform specialized radiologists and their potential role in identifying affected patients as soon as possible. I shared it.

“We have successfully developed a deep learning algorithm for estimating the malignant risk of lung nodules detected on low-dose screening CT, which can be generalized across screening populations and protocols,” said a doctoral candidate for diagnostic imaging. Said the team led by the lead author, Kiran Vaidhya Venkadesh. Analytical group at Radboud University Medical Center. “This deep learning algorithm can help radiologists optimize follow-up recommendations for participants undergoing lung cancer screening, which can help reduce unnecessary diagnostic interventions. “

It can also reduce the burden on radiologists and reduce the cost of lung cancer screening.

To determine the performance of the algorithm, the team used the results as an established and effective PanCan lung cancer early detection model, and 11 clinicians (4 thoracic radiologists, 5 radiations). Compared with the results of a clinician and two pulmonologists).

In a retrospective study, we developed an algorithm using deep learning (available for free) Here) And it was trained on CT images of 16,077 nodules containing 1,249 malignant tumors. Images were collected from the National Lung Screening Trial between 2002 and 2004. They tested the algorithm using three sets of imaging data from the Danish lung cancer screening trial: a complete cohort of all 883 nodules (65 malignant), and two size-matched. Rich cohort (175 nodules, 59 malignant) Size matching (177 nodules, 59 malignant).

Relevant content: Artificial intelligence chest x-ray catches more lung cancer

When Venkadesh’s team compared the performance of the algorithm to both existing evaluation models and clinician performance, they found that the algorithm far outperformed both. For the PanCan model, the algorithm achieved an area under the curve (AUC) of 0.93 compared to 0.90.

“The algorithm significantly outperformed the PanCan model only in a cancer-rich subset of size matching,” the team explained. “This suggests that the algorithm is more dependent on the discriminating imaging properties than the PanCan model, even though nodule size remains a strong predictor of malignancies.”

Relevant content: Added Predicted Heart Disease Death Risk to LDCT Function

Both random benign nodules (AUC 0.96 vs 0.90) and size-matched benign nodules (AUC 0.86 vs 0.82) outperformed the cancer-rich cohort of thoracic radiologists.

Based on these results, the team stated that the algorithm could provide some benefits to the clinical environment. Radiologists can upgrade suspicious nodules to the Lung-RADS4X category, but the algorithm does not require manual interpretation of the image properties of the nodules. This can lead to a substantial reduction in observational variability in CT interpretation, said senior author Dr. Colin Jacobs, an assistant professor of medical imaging at Radboud.

Ultimately, the team said they believe the algorithm is being used as a support tool for the efforts of radiologists.

“With the help of a reliable artificial intelligence system, we anticipate the demand for trained human observers who will act as the first leaders in chest CT when lung cancer screening programs are introduced worldwide. “I will,” said the team. “This deep learning algorithm can help radiologists optimize follow-up recommendations for participants undergoing lung cancer screening, which can help reduce unnecessary diagnostic interventions. “

In an accompanying editorial, PanCan developers Martin C. Tammemägi, DVM, MSc, Ph.D. Reiterated the need for algorithms that help distinguish between malignant and benign nodules and reduce the workload of the provider. He said Venkadesh’s team has improved past AI prediction models with promising results. However, he warned not to overemphasize the achievements of AUC.

“I warn readers not to over-interpret AUC. In many cases, AUC is directly interpreted as a measure of predictive accuracy,” he explained. “AUC is not a percentage … AUC does not measure absolute classification accuracy, but it evaluates whether the model can place uppercase and lowercase pairs in the correct rank order.”

This makes algorithm calibration important, he said. Also, given that the algorithm misidentified a malignant nodule as benign and vice versa, calibration problems could almost certainly occur.

“The clinician’s judgment [deep learning] Extreme and inaccurate scores of the algorithm … it is possible to imagine that harm could be done, “he said.

Still, he said, the relatively high AUC achieved by the algorithm indicates that it is getting valuable predictive information from factors other than size.

But according to the Ben Kaedesh team, their work is not perfect. They are currently working on another algorithm that uses multiple CT examinations at input, from initial or baseline screening to subsequent screening to help compare nodule growth and appearance to previous scans. May expand to.

For more information based on industry expert insights and research, subscribe to our diagnostic imaging newsletter. Here..