Deep learning tools improve the accuracy of malignant detection of pulmonary nodules

Artificial Intelligence (AI) deep learning tools to estimate malignant risk of pulmonary nodules have significantly reduced false positive results while also providing high cancer detection rates. The results of a study using data from a large multisite lung cancer screening trial were published today. RadiologyJournal of the North American Radiology Society (RSNA).

Lung cancer remains a major global health problem and is causing the world's most cancer-related deaths. Screening of high-risk individuals with low-dose chest CT has been shown to reduce lung cancer mortality. However, early screening trials reported high false positive rates, leading to unnecessary follow-up procedures, patient anxiety and increased healthcare costs.

Small round or oval growth of lung nodules of lungs is common, and malignant identification is difficult in screening for lung cancer.

Deep learning offers promising solutions, but robust verification is essential. AI further evaluates the potential for making nodules even malignant, explaining factors that don't even appear on CT scans. ”

NOA Antonissen, MD, Lead Researcher and PhD Candidate for Radboud University Medical Center, Nimegen, Netherlands

Most current lung cancer screening protocols rely on nodule size, type, and growth to estimate malignant risk. Early detection of pan-Canadians in a lung cancer (Pancan) model that estimates nodal malignant risk through a combination of patient-nodal characteristics demonstrates how probability-based tools can improve risk assessment. Such probability-based risk thresholds are increasingly being used to guide management protocols. Deep learning offers a promising alternative by enabling fully data-driven predictions, but requires more evidence before adoption in clinical practice.

In a retrospective study, researchers trained in-house development deep learning algorithms to estimate the risk of pulmonary nodule malignant tumors using data from the National Lung Screening Trial, which includes 16,077 nodules (1,249 malignant).

External studies were conducted using baseline CT scans of the Danish lung cancer screening trial, multiventricular-centered Italian lung detection trial, and the Netherlands and Belgia Nelson trial. The pooled cohort included 4,146 participants (median 58 years, 78% male, median smoking history of 38 years), with 7,614 benign and 180 malignant nodules.

The researchers evaluated the performance of the pooled cohort and two subsets of algorithms. It is an indeterminate nodule (5-15 mm) and a malignant nodule that is in size consistent with the benign nodule.

“Due to diagnostic challenges and frequent needs for short-term follow-up, we chose nodules that were 5-15 mm in size,” said Dr. Antonissen. “An accurate risk classification of these nodules may reduce unnecessary procedures.”

For comparison, the performance of the algorithm was assessed against nodal and participant level Pancan models using the area under the receiver operating characteristic curve (AUC). The AUC summarizes whether the model can generate relative scores to distinguish between positive or negative instances across all classification thresholds.

In the pooled cohort, the deep learning model achieved AUCS of 0.98, 0.96, and 0.94 for cancers diagnosed within 1 and 2 years, respectively, compared to Pancans of 0.98, 0.94, and 0.93.

For indeterminate nodules (129 malignant, 2,086 benign), the deep learning model significantly outperformed Pankan in all time frames with AUCS of 0.95, 0.94, 0.90 vs 0.91, 0.88, and 0.86. For cancers that matched the benign nodule size (180 malignant, 360 benign), the AUC in the deep learning model was 0.60 and 0.79 for Pancan.

With 100% sensitivity of cancer diagnosed within one year, the deep learning model classifies 68.1% as low risk compared to 47.4% using the PANCAN model, representing a relative reduction of 39.4% for false positives.

“Deep learning algorithms can help radiologists determine whether follow-up imaging is necessary, but future verification is needed to determine the clinical applicability of these tools and actually guide their implementation,” said Dr. Antonissen. “Reducing false positive results makes screening for lung cancer more feasible.”