Identifying transcription patterns of histology and radiomics features in NSCLC with neural networks

Abstract

AbstractPurposeTo investigate the use of deep neural networks to learn associations between gene expression and radiomics or histology in non-small cell lung cancer (NSCLC).Materials and MethodsDeep feedforward neural networks were used for radio-genomic mapping, where 21,766 gene expressions were inputs to individually predict histology and 101 CT radiomic features. Models were compared against logistic regression, support vector machines, random forests, and gradient boosted trees on 262 training and 89 testing patients. Neural networks were interpreted using gene masking to derive the learned associations between subsets of gene expressions to a radiomic feature or histology type.ResultsNeural networks outperformed other classifiers except in five radiomic features, where training differences were <0.026 AUC. In testing, neural networks classified histology with AUCs of 0.86 (adenocarcinoma), 0.91 (squamous), and 0.71 (other); and 14 radiomics features with >= 0.70 AUC. Gene masking of the models showed new and previously reported histology-gene or radiogenomic associations. For example, hypoxia genes could predict histology with >0.90 test AUC and published gene signatures for histology prediction were also predictive in our models (>0.80 test AUC). Gene sets related to the immune or cardiac systems and cell development processes were predictive (>0.70 test AUC) of several different radiomic features while AKT signaling, TNF, and Rho gene sets were each predictive of tumor textures.ConclusionWe demonstrate the ability of neural networks to map gene expressions to radiomic features and histology in NSCLC and interpret the models to identify predictive genes associated with each feature or type.Author SummaryNon-small-cell lung cancer (NSCLC) patients can have different presentations as seen in the CT scans, tumor gene expressions, or histology types. To improve the understanding of these complementary data types, this study attempts to map tumor gene expressions associated with a patient’s CT radiomic features or a histology type. We explore a deep neural network approach to learn gene-radiomic associations (i.e., the subsets of co-expressed genes that are predictive of a value of an individual radiomic feature) and gene-histology associations in two separate public cohorts. Our modeling approach is capable of learning relevant information by showing the model can predict histology and that the learned relationships are consistent with prior works. The study provides evidence for coherent patterns between gene expressions and radiomic features and suggests such integrated associations could improve patient stratification.