Ensemble framework based on attributes and deep features for benign-malignant classification of lung nodule

Abstract

Early detection and identification of malignant lung nodules improve the survival of lung cancer patients. The visual attributes such as subtlety, spiculation, and calcification of lung nodules play an important role in the diagnosis of malignancy. However, the gap between attributes and computation features is the main factor that restricts the performance of computer-aided diagnosis (CAD). Therefore, we propose a Fuse-Long Short-Term Memory-Convolutional Neural Network (F-LSTM-CNN) ensemble learning algorithm which incorporates visual attributes and deep features to classify benign and malignant nodules. First, the attribute features are obtained from clinical information while the deep features of nodules are extracted from the preprocessed computed tomography (CT) images. Second, the Fuse-Convolutional Neural Network (F-CNN) model is proposed for highlighting the essential role of attributes in the classification processing which integrates deep features and attribute features mapped through the transposed convolution. Meanwhile, the Fuse-Long Short-Term Memory (F-LSTM) model is proposed to focus on the specific deep features for classification via the affine transformation of attribute features. Finally, early identification of malignant lung nodules is conducted by fusing the prediction scores of the F-LSTM and F-CNN models. The experiments were conducted on the public lung nodule dataset (LIDC-IDRI) and achieved accuracy, sensitivity, and specificity of 0.955, 1, and 0.937 with an Area under the ROC Curve (AUC) of 0.995 for lung nodule classification. The experiment results show that the proposed F-LSTM-CNN ensemble learning model facilitates the interpretation of diagnostic data and helps radiologists to make decisions in clinical practice.