3D Multi-scale DenseNet for Malignancy Grade Classification of Pulmonary Nodules

Abstract

With the recent development of algorithm for computer-aided diagnosis (CAD) system, detection of pulmonary nodules from computed tomography (CT) imaging data with high accuracy is becoming possible. Existing CAD system is able to automatically output the location of a nodule with its confidence. It helps the radiologist to save time for nodule detection work. However, not all nodules will develop into lung cancer. Depending on its grade of malignancy, the probability of developing into lung cancer is different. Evaluating the grade of malignancy of pulmonary nodule is performed by doctors and highly depends on personal experience. In order to further automate the process of lung cancer prognosis, a system that accurately evaluates the grade of malignancy of a pulmonary nodule is needed. It will be helpful to re-evaluate the detected nodules and provide proper suggestion for therapeutic method. There are two types of tasks for malignancy classification (1) to classify a sample into benign or malignant (2) to classify a sample into malignancy grades (from grade-1 to grade-5). Many researches have achieved a high accuracy for task-1, but the results on task-2 are still poor. In this work, we present a 3D Multi-scale DenseNet to classify the grade of malignancy of pulmonary nodules. Through the observation of CT image data, we found that for some small nodules it is impossible to extract their morphological features due to their small size. Our idea is to convert the original CT image into three different scales (Multi-scale) and input them into three parallel 3D densely-connected convolutional network (DenseN et) blocks. Finally, the extracted features from the last layer of the three networks are concatenated to classify the grade of malignancy. In this way, the morphological features of small nodules can be better obtained without affecting the feature extraction of large nodules. In this study, 1882 samples from the dataset of Lung Image Database Consortium (LID C) are used for training and testing. Overall, we achieved 68.5 % accuracy for the task of malignancy grades classification.