Using deep-learning radiomics to predict lung cancer histology.

8545 Background: Histologic phenotype is an important predictor of clinical outcomes in lung cancer. Tissue diagnosis is the most definitive approach to categorization, however, this is often technically challenging for thoracic lesions. In this study, we explore Deep-Learning Radiomics methods for non-invasive histology classification in early-stage Non-Small Cell Lung Cancer (NSCLC). Methods: A cohort of 157 patients with Stage I NSCLC identified as either adenocarcinoma or squamous cell carcinoma on pathology was used. All patients were surgical candidates at Massachusetts General Hospital between 2004-2010. Deep feature extraction from pretreatment CT images was conducted using a pre-trained VGG-16 convolutional neural network (CNN). In addition to appending fully-connected classifying layers to the network, a transfer learning approach was also employed using different classifiers. Three machine-learning classification models were independently evaluated on the extracted features: K-Nearest Neighbors (kNN), Random Forest Classifier (RF), and Least Absolute Shrinkage and Selection Operator (LASSO). Principal component analysis was employed in selecting features corresponding to 90% cumulative explained variance. A LASSO method was then used to select the best features. Models were trained on 100 patients and cross-validated on an independent test-set of 57 patients. Results: All models were able to perform binary classification of tumor histology (adenocarcinoma vs squamous cell carcinoma). The fully-connected CNN had the highest performance (AUC = 0.751). Other classifiers also showed significant predictive power after dimension reduction of the feature space (from 512 to 46), with AUC = 0.712 for LASSO (α = 0.1), and AUC = 0.689 for kNN (k = 5). RF had the lowest predictive performance (AUC = 0.533). 73% of the study group had adenocarcinoma vs 27% with squamous cell carcinoma, and ratios were balanced between training and validation sets. Conclusions: Deep-Learning Radiomics is a promising approach to non-invasive lung cancer histology classification. These methods can potentially augment other emerging techniques, such as liquid biopsy; offering complementary information to help in clinical decision making.