Deep learning-based automated Lung-RADS classification algorithm for LDCT lung cancer screening

Purpose: The Lung Imaging Reporting and Data System (Lung-RADS) criteria are widely adopted for global lung cancer screening [1]. Nevertheless, the task is not only time-consuming due to the interpretation of hundreds of LDCT slices but also exhibits variability in classification among radiologists [2]. To address these challenges, we have devised a deep learning-based automated Lung-RADS classification algorithm.

Methods/Materials: We collected 1,785 chest LDCT scans obtained at Pusan National University Hospital between 2019 and 2022 for the development of the proposed algorithm. The modified NoduleNet [3] was trained to detect the location of lung nodules with transfer learning. The patches used as input to the subsequent segmentation model were obtained using the bounding box of the detected nodules. A segmentation model based on 3D Unet3+ was used to classify the texture of detected nodules into 3-class (solid, ground-glass-opacity, and part-solid) and then extracted masks, and a classification-guided module was included to filter the extracted masks for the overall model performance improvement. The size of the nodule was measured through the average length of the major and minor axes in the central slice of the nodule. Our model reflected the calculation of the Lung-RADS score by radiologists, so the Lung-RADS considered the size and texture of nodules based on the Lung-RADS version 1.1. In addition, for specific performance comparison and analysis, only nodules with a Dice score of 0.2 or more were evaluated [4], so that localization of nodules was considered. For the evaluation of nodule detection, we employed sensitivity and specificity metrics. Subsequently, the overall performance of our system was assessed using the accuracy of the Lung-RADS score.

Results: The performance of the nodule detection model increased by 1.38 and 1.13 times in sensitivity and specificity, respectively, compared to the conventional NoduleNet. As a result of the LUNG-RADS classification, the accuracy of each Lung-RADS 2, 3, 4A, and 4B was measured as 61.02, 78.45, 81.41, and 96.38 %.

Conclusions: Sufficient diagnostic accuracy was obtained through the proposed two-stage model and the provided Lung-RADS score.