Myxomavirus (MYXV) therapy for small cell lung cancer (SCLC) using patient samples and a genetically engineered SCLC mouse model

Abstract

between benign and cancerous nodules, and ii) quantitatively predict risk of lung cancer incidence. Methods: Using data and images from the NLST, we performed post hoc nested case-control analyses. The first analysis was conducted to identify diagnostic quantitative imaging features that differentiate between malignant tumors and benign nodules. This study included 88 incidence lung cancer cases diagnosed at the first follow-up interval (T1) and 172 “controls” that had a noduleþ scan at T1 that was not lung cancer. The second analysis was conducted to identify predictive quantitative imaging features that are predictive of lung cancer risk. This study utilized baseline scans (T0) from 74 subjects who developed an incidence lung cancer in follow-up intervals and 125 “controls” that had a noduleþ result in follow-up intervals that was not lung cancer. The LDCT scans were subjected to an in-house “Radiomic Pipeline” that converts images to mineable data (>400 quantitative features). Two classes of features were extracted: semantic and agnostic. Semantic features are commonly used in the radiology lexicon to describe regions of interest. Agnostic features are mathematically extracted quantitative descriptors that capture lesion heterogeneity. Separate statistical analyses were performed for the diagnostic and predictive features. Univariable analyses and false discovery rate (FDR) were utilized to identify which were features were statistically significant. To generate a parsimonious model, we performed a backward elimination process using a 0.1 threshold for inclusion. Results: Although nodule size has diagnostic utility, especially among the largest nodules, >80% of cases and controls had nodules <15 cm. For size alone, we found a modest AUC of 0.79 when nodules were <15 cm. We sought to improve the diagnostic capability of size by adding imaging features. Univariable analyses revealed that 17 of the features were significantly different between cases and controls. Backward elimination process revealed a model with 3 imaging features (radius of smallest enclosing ellipse, radius of largest enclosed ellipse, and tumor thickness-pixel) that yielded an AUC of 0.88; and a model with those 3 features, size, and demographics yields an AUC of 0.89. For the risk prediction analysis, univariable analyses revealed that 10 of the features were significantly associated with lung cancer risk which remained significant when included in a single model including demographics/risk factors. Backward elimination process identified a model with six imaging features (concavity, border definition, attachment to vessel, perinodule emphysema, perinodule fibrosis, nodules in both lungs) and demographics yielding an AUC of 0.87 compared to 0.58 for demographics alone. Conclusions: These results demonstrate that we can improve the diagnostic utility of size alone by including additional imaging features. Moreover, these data provide strong and compelling evidence for the utility of imaging features for risk prediction.