Benign/Malignant Lung Nodule Classification Using Inter-label Dependency

Lung cancer has the highest incidence and mortality in China.Early screening of ma-lignant lung nodules can reduce the mortality of lung cancer remarkably.Chest computed to-mography(CT)images can capture lung nodule heterogeneity in a non-invasive way,which has become the most common approach for early screening of lung cancer.Precise identifi-cation of the nature of lung nodules is extremely difficult,not only because clinic doctors can not maintain a high concentration to make accurate diagnosis under high-intensity workloads,but also because human eyes are difficult to capture the lung nodule heterogeneity that hap-pens at subtle changes in CT values.These problems have been alleviated with the advent of computer-assisted diagnosis(CAD)systems for lung cancer,which can automatically diagnose lung nodules in chest CT images.Recently,the application of deep convolutional neural net-works(DCNNs)in lung nodule detection and classification makes lung cancer CAD systems achieve promising progress.However,compared with natural image classification,there exist small data and category imbalance issues in the benign/malignant lung nodule classification task.It is necessary to extract lung nodule heterogeneity features from limited and imbalanced training data when using DCNN.In this paper,we focus on classifying the nature of lung nodules using deep learning tech-niques,and propose a novel benign/malignant lung nodule classification method based on the inter-label dependency.The proposed method effectively address the small data and the multi-label category imbalance issues in this classification task.The main contributions of this paper include:1.The method establishing the inter-label dependency through multi-label classification is proposed,enabling the model to discriminate the nature of lung nodules by leveraging their radiographic attributes,thereby addressing the small data issue on the benign/malignant lung nodule classification task.2.A multi-label softmax loss function is proposed to directly optimize the ranking loss between the labels and the overall AUC metric.Such metrics can better evalute the classification performance of the model on the multi-label category imbalanced dataset.3.A novel component is introduced for the proposed loss function,which can control the overwhelming gradients between the labels.The component improves the overall performance of the multi-label classifier.4.A pathological-based benign/malignant lung nodule dataset and a practically usable intelligent diagnosis system for lung nodules were constructed.The experimental results on this dataset verify the effectiveness of the proposed method,and the model trained on this dataset is applied in the intelligent diagnosis system.