Abstract
BACKGROUNDContinued improvement in deep learning methodologies has increased the rate at which deep neural networks are being evaluated for medical applications, including diagnosis of lung cancer. However, there has been limited exploration of the underlying radiological characteristics that the network relies on to identify lung cancer in computed tomography (CT) images.OBJECTIVEIn this study, we used a combination of image masking and saliency activation maps to systematically explore the contributions of both parenchymal and tumor regions in a CT image to the classification of indeterminate lung nodules.METHODSWe selected individuals from the National Lung Screening Trial (NLST) with solid pulmonary nodules 4-20 mm in diameter. Segmentation masks were used to generate three distinct datasets; 1) an Original Dataset containing the complete low-dose CT scans from the NLST, 2) a Parenchyma-Only Dataset in which the tumor regions were covered by a mask, and 3) a Tumor-Only Dataset in which only the tumor regions were included.RESULTSThe Original Dataset significantly outperformed the Parenchyma-Only Dataset and the Tumor-Only Dataset with an AUC of 80.80 ± 3.77% compared to 76.39 ± 3.16% and 78.11 ± 4.32%, respectively. Gradient-weighted class activation mapping (Grad-CAM) of the Original Dataset showed increased attention was being given to the nodule and the tumor-parenchyma boundary when nodules were classified as malignant. This pattern of attention remained unchanged in the case of the Parenchyma-Only Dataset. Nodule size and first-order statistical features of the nodules were significantly different with the average malignant and benign nodule maximum 3d diameter being 23 mm and 12 mm, respectively.CONCLUSIONWe conclude that network performance is linked to textural features of nodules such as kurtosis, entropy and intensity, as well as morphological features such as sphericity and diameter. Furthermore, textural features are more positively associated with malignancy than morphological features.