Detectability of pulmonary nodules by deep learning: results from a phantom study

Abstract

To investigate how nodule size, nodule density, scan dose, slice thickness and reconstruction methods affect the performance of a deep learning (DL) model for detection of pulmonary nodules in phantom CT scans. Spherical lung nodule phantoms of two different densities (− 630 HU and + 100 HU) and five different sizes (3, 5, 8, 10, and 12 mm) were inserted into an anthropomorphic chest phantom. CT data were scanned and reconstructed using three different tube current (10, 50, 200 mAs), two different slice thickness of 1 and 2 mm, four reconstruction methods (FBP-standard (FBP-STD), FBP-Y sharp kernel (FBP-YA), iDose4-standard kernel (iDose4-STD) and iDose4-Y sharp kernel (iDose4-YA). Evaluation of deep learning model focused on detection sensitivity and precision. According to the statistical results from the study, we found that the sensitivity and precision performance depends on the nodule sizes, nodule type, tube current, reconstruction methods and image thickness. Comparing the solid (100 HU) and ground-glass (GGO, − 630 HU) nodule phantoms, solid nodule phantom predictions are rarely affected by tube current, reconstruction methods and nodule sizes. Both sensitivity and precision are close to 100% in all solid nodule phantom prediction cases. While the sensitivity and precision metrics of GGO nodule phantoms change in a wide range from 42.9 to 100%. Larger nodule size and higher tube current gives a better sensitivity and precision for GGO nodule phantoms in most cases. We also analyze the relationships between the image thickness and the reconstruction methods. For 1-mm thickness images, iDose4-STD and FBP-STD shows a better result in both sensitivity and precision metrics. As for 2-mm thickness images, iDose-YA and FBP-YA gain a better performance. The results of this phantom study demonstrated that high stability and flexibility of deep learning model can be used in daily clinical and screening practice.