Perinodular and Intranodular Radiomic Features on Lung CT Images Distinguish Adenocarcinomas from Granulomas

Abstract

Purpose To evaluate ability of radiomic (computer-extracted imaging) features to distinguish non-small cell lung cancer adenocarcinomas from granulomas at noncontrast CT. Materials and Methods For this retrospective study, screening or standard diagnostic noncontrast CT images were collected for 290 patients (mean age, 68 years; range, 18-92 years; 125 men [mean age, 67 years; range, 18-90 years] and 165 women [mean age, 68 years; range, 33-92 years]) from two institutions between 2007 and 2013. Histopathologic analysis was available for one nodule per patient. Corresponding nodule of interest was identified on axial CT images by a radiologist with manual annotation. Nodule shape, wavelet (Gabor), and texture-based (Haralick and Laws energy) features were extracted from intra- and perinodular regions. Features were pruned to train machine learning classifiers with 145 patients. In a test set of 145 patients, classifier results were compared against a convolutional neural network (CNN) and diagnostic readings of two radiologists. Results Support vector machine classifier with intranodular radiomic features achieved an area under the receiver operating characteristic curve (AUC) of 0.75 on the test set. Combining radiomics of intranodular with perinodular regions improved the AUC to 0.80. On the same test set, CNN resulted in an AUC of 0.76. Radiologist readers achieved AUCs of 0.61 and 0.60, respectively. Conclusion Radiomic features from intranodular and perinodular regions of nodules can distinguish non-small cell lung cancer adenocarcinomas from benign granulomas at noncontrast CT. © RSNA, 2018 Online supplemental material is available for this article. See also the editorial by Nishino in this issue.

Figure 1:

A, Images show overview of methodology. CT images were retrospectively collected. Region of interest was manually segmented in axial view to obtain intranodular mask, and perinodular masks were automatically generated for varying distances (shown here at 5 mm) outside tumor. Haralick, Laws energy, Gabor texture, and shape features were extracted from largest tumor slice. Next, t test was implemented to select top 12 features to train support vector machine classifier and validate it on independent set (n = 145). B, Diagram shows features extracted in each experiment before feature selection was implemented. Total of 1776 features were extracted from each solitary pulmonary nodule, with 252 features from intranodular region, 12 shape features, and 1512 features from perinodular regions (252 features from each of the six perinodular regions).

Figure 2:

Consolidated Standards of Reporting Trials, or CONSORT, flow diagram of patient enrollment, eligibility, and exclusion criteria of data set.

Figure 3:

Images show feature extraction from perinodular region. A, B, Noncontrast CT scans in axial view of adenocarcinoma in a 55-year-old man and granuloma in a 67-year-old woman, respectively. C, Heat map of Haralick entropy feature that was extracted from lung parenchyma (also termed perinodular region of lung nodule) demonstrates, D, various intervals (up to 30 mm of lung parenchyma, outside lung nodule) from which radiomic features were extracted as annular rings.

Figure 4:

Images show intranodular and perinodular radiomics of adenocarcinoma in a 61-year-old woman at noncontrast CT (axial view). A, Higher spatial resolution image of malignant nodule. B, Top row represents higher expression of Laws energy (E5W5) inside tumor; E5W5 implies that Laws energy–based textural patterns of edges (or E ) in horizontal direction and waves (or W ) in vertical direction by using five-pixel by five-pixel two-dimensional convolution filter. Further information on Laws energy features can be found in Table E2 (online). Bottom row shows representative hematoxylin and eosin (H&E) stain of adenocarcinoma (original magnification, ×100), where tumor cells show high nucleus-to-cytoplasm ratio with irregularly shaped nuclei. Tumor cells form angulated irregular acini in background of fibrosis. Scant intranodular lymphocytes are present. Pigmented macrophages are present within malignant acini. C, Top row represents radiomic heat map of low-frequency Gabor feature, which is expressed higher in adenocarcinomas in peritumoral region, and bottom row is H&E stain of adenocarcinoma (original magnification, ×100), where increased lymphocytes and macrophages are observed at interface between tumor and normal lung. This “rim” of lymphocytes and macrophages is less than 1 mm. D, Shape of entire malignant nodule.

Figure 5:

Images show intranodular and perinodular radiomics of granuloma in a 55-year-old man at noncontrast CT (axial view). A, Higher spatial resolution image of benign nodule. B, Top row represents lower expression of Laws energy (E5 W5) inside tumor; E5W5 implies that Laws energy-based textural patterns of edges (or E ) in horizontal direction and waves (or W ) in vertical direction by using five-pixel by five-pixel two-dimensional convolution filter. Further information on Laws energy features can be found in Table E2 (online). Bottom row shows representative hematoxylin and eosin (H&E) stain of granuloma (original magnification, ×100), where it consists of admixed lymphocytes, plasma cells, and histiocytes. There are also areas of fibrosis, necrosis, and calcification. C, Top row represents radiomic heat map of low-frequency Gabor feature, which has low expression in granulomas in peritumoral region, and bottom row is H&E stain of adenocarcinoma (original magnification, ×100), where giant cells are observed at interface between nodule and normal lung. D, Shape of entire benign nodule.

Figure 6:

A, Graph shows unsupervised hierarchal clustering of intranodular and perinodular radiomic features. The x-axis represents top 12 features where (P) denotes a perinodular feature; y-axis represents independent test set of patients (n = 145). Dendrogram highlighted in red represents prominent cluster of adenocarcinomas. B, First row shows noncontrast baseline lung CT scans (axial view) of granuloma nodule in an 81-year-old man from independent test set, which was diagnosed as “mostly malignant” by both expert readers (score of 4). Radiomic heat map represents Laws energy feature inside nodule and also low expression of high-frequency Gabor response captured in perinodular region of 0–5 mm outside nodule. Second row shows noncontrast baseline lung CT scans (axial view) from independent test set of adenocarcinoma nodule in a 63-year-old woman from independent test set, which was diagnosed as a “mostly benign” granuloma by expert reader 1 (score of 2) and “not sure” by expert reader 2 (score of 3). Radiomic heat map represents Laws energy feature inside nodule and also high expression of low-frequency Gabor response (f = 8) captured in perinodular region of 0–5 mm outside nodule. These cases were correctly classified by linear support vector machine classifier that was trained.