Nodule type classification for lung cancer screening with CT

Doohyun Park, Jinyoung Kim

Abstract:

Introduction:
The American College of Radiology has periodically updated its lung cancer screening guidelines, known as the Lung Imaging Reporting and Data System (Lung-RADS). Despite these updates, distinguishing nodule types—solid, part-solid, and ground-glass nodules (GGN)—remains clinically important. For instance, patient management guideline varies depending on nodule type, even when nodules have the same size. Therefore, precise classification of lung nodule types lays the groundwork for nodule management. Typically, radiologists assess by comparing adjacent pulmonary vessels, because ground-glass opacity refers to an area of increased attenuation that does not completely obscure the underlying bronchial and vascular structures. However, this method is not objective and can lead to observer variability. This variability emphasizes the need for accurate and automated methods to classify nodule types. In response, this study employs commercial software (VUNO Med-LungCT AI) to evaluate its capability in classifying lung nodule types.
Methods:
We collected data on 238 patients with nodules from a referral hospital in the United States. Two radiologists reviewed the images and identified 1,896 nodules by consensus. From these, 11 part-solid nodules were excluded. Out of the remaining 1,885 nodules, 1,413 (75%) were used to update the model, and 472 (25%) were used for testing. The test set included 422 (89%) solid nodules and 50 (11%) GGNs. This study used VUNO Med-LungCT AI for nodule type classification.
Result:
The classification of 472 nodules yielded an area under the receiver operating characteristic (ROC) curve (AUC) of 0.972 (95% confidence interval [CI], 0.953–0.991). The accuracy was 98.3% for solid nodules (415 out of 422) and 86.0% for GGNs (43 out of 50). Figure 1.A shows the ROC curve, and Figure 1.B displays axial slice CT images from eight cases, incorrectly classified as the opposite type. The qualitative analysis indicates that most of the incorrectly classified nodules were difficult to categorize definitively even by clinicians, because they are small or have ambiguous HU values. This ambiguity means that while some clinicians might classify such nodules as solid, others could identify them as GGN. The classification of these 472 nodules was completed in under one second, demonstrating the software's efficiency in rapid processing.
Conclusion:
VUNO Med-LungCT AI demonstrated impressive performance in nodule type classification. Future studies will expand the part-solid and cystic nodules according to the Lung-RADS 2022. With these enhancements, the VUNO Med-LungCT AI is expected to achieve superior performance in Lung-RADS categorization.