LDCT image biomarkers that matter most for the deep learning classification of indeterminate pulmonary nodules.

BACKGROUND: Continued 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. OBJECTIVE: In 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. METHODS: We 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. RESULTS: The 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. CONCLUSION: We 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.

Perinodular Parenchymal Features Improve Indeterminate Lung Nodule Classification.

BACKGROUND: Radiomics, defined as quantitative features extracted from images, provide a non-invasive means of assessing malignant versus benign pulmonary nodules. In this study, we evaluate the consistency with which perinodular radiomics extracted from low-dose computed tomography images serve to identify malignant pulmonary nodules. MATERIALS AND METHODS: Using the National Lung Screening Trial (NLST), we selected individuals with pulmonary nodules between 4mm to 20mm in diameter. Nodules were segmented to generate four distinct datasets; 1) a Tumor dataset containing tumor-specific features, 2) a 10 mm Band dataset containing parenchymal features between the segmented nodule boundary and 10mm out from the boundary, 3) a 15mm Band dataset, and 4) a Tumor Size dataset containing the maximum nodule diameter. Models to predict malignancy were constructed using support-vector machine (SVM), random forest (RF), and least absolute shrinkage and selection operator (LASSO) approaches. Ten-fold cross validation with 10 repetitions per fold was used to evaluate the performance of each approach applied to each dataset. RESULTS: With respect to the RF, the Tumor, 10mm Band, and 15mm Band datasets achieved areas under the receiver-operator curve (AUC) of 84.44%, 84.09%, and 81.57%, respectively. Significant differences in performance were observed between the Tumor and 15mm Band datasets (adj. p-value <0.001). However, when combining tumor-specific features with perinodular features, the 10mm Band + Tumor and 15mm Band + Tumor datasets (AUC 87.87% and 86.75%, respectively) performed significantly better than the Tumor Size dataset (66.76%) or the Tumor dataset. Similarly, the AUCs from the SVM and LASSO were 84.71% and 88.91%, respectively, for the 10mm Band + Tumor. CONCLUSIONS: The combined 10mm Band + Tumor dataset improved the differentiation between benign and malignant lung nodules compared to the Tumor datasets across all methodologies. This demonstrates that parenchymal features capture novel diagnostic information beyond that present in the nodule itself. (data agreement: NLST-163).

Wavelet decomposition facilitates training on small datasets for medical image classification by deep learning.

The adoption of low-dose computed tomography (LDCT) as the standard of care for lung cancer screening results in decreased mortality rates in high-risk population while increasing false-positive rate. Convolutional neural networks provide an ideal opportunity to improve malignant nodule detection; however, due to the lack of large adjudicated medical datasets these networks suffer from poor generalizability and overfitting. Using computed tomography images of the thorax from the National Lung Screening Trial (NLST), we compared discrete wavelet transforms (DWTs) against convolutional layers found in a CNN in order to evaluate their ability to classify suspicious lung nodules as either malignant or benign. We explored the use of the DWT as an alternative to the convolutional operations within CNNs in order to decrease the number of parameters to be estimated during training and reduce the risk of overfitting. We found that multi-level DWT performed better than convolutional layers when multiple kernel resolutions were utilized, yielding areas under the receiver-operating curve (AUC) of 94% and 92%, respectively. Furthermore, we found that multi-level DWT reduced the number of network parameters requiring evaluation when compared to a CNN and had a substantially faster convergence rate. We conclude that utilizing multi-level DWT composition in place of early convolutional layers within a DNN may improve for image classification in data-limited domains.