Room: Exhibit Hall | Forum 8
Purpose: In this work we present Deep-Learning based Radiomics models for non-invasive histology classification in Non-Small Cell Lung Cancer (NSCLC) using CT images.
Methods: A cohort of 157 patients surgically treated for Stage-I NSCLC at Massachusetts General Hospital was used. All patients were identified on pathology as having either Adenocarcinoma or Squamous Cell Carcinoma. A transfer-learning approach was used for image based histologic subtype classification. A pre-trained VGG-16 neural network was used as a fixed feature extractor for pretreatment CT images. 512-D feature vectors were computed for each image. Dimensionality reduction of the deep radiomic feature space was conducted using Principal Component Analysis. The LASSO method was then used to select the best features. Three Machine-Learning classifiers were independently evaluated on the features: kNN, RF, and LASSO. In addition, fully-connected classifying layers were also appended to the VGG-16 network. 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 neural network had the highest performance (AUC = 0.751). Other classifiers also showed significant predictive power after dimension reduction of the feature space to 46-D, with AUC = 0.712 for LASSO, and AUC = 0.689 for kNN. RF had the lowest predictive performance (AUC = 0.533). 73% of the study group had Adenocarcinoma vs 27% with Squamous Cell Carcinoma, and event rates were balanced between training and validation sets.
Conclusion: Deep-Learning Radiomics is a promising approach to non-invasive lung cancer histology classification, and has the potential to advance precision medicine.