RESUMO
The accurate identification of malignant lung nodules using computed tomography (CT) screening images is vital for the early detection of lung cancer. It also offers patients the best chance of cure, because non-invasive CT imaging has the ability to capture intra-tumoral heterogeneity. Deep learning methods have obtained promising results for the malignancy identification problem; however, two substantial challenges still remain. First, small datasets cannot insufficiently train the model and tend to overfit it. Second, category imbalance in the data is a problem. In this paper, we propose a method called MSCS-DeepLN that evaluates lung nodule malignancy and simultaneously solves these two problems. Three light models are trained and combined to evaluate the malignancy of a lung nodule. Three-dimensional convolutional neural networks (CNNs) are employed as the backbone of each light model to extract the lung nodule features from CT images and preserve lung nodule spatial heterogeneity. Multi-scale input cropped from CT images enables the sub-networks to learn the multi-level contextual features and preserve diverse. To tackle the imbalance problem, our proposed method employs an AUC approximation as the penalty term. During training, the error in this penalty term is generated from each major and minor class pair, so that negatives and positives can contribute equally to updating this model. Based on these methods, we obtain state-of-the-art results on the LIDC-IDRI dataset. Furthermore, we constructed a new dataset collected from a grade-A tertiary hospital and annotated using biopsy-based cytological analysis to verify the performance of our method in clinical practice.