LA-Net: Lung Adenocarcinoma Classification with Assistants from Lung Nodule Classification and Positional Information

Abstract

Lung cancer (LC) is one of the most common and fatal cancer in human. Recent LC diagnostic studies focus on using convolutional neural networks (CNNs) and achieve certain success. However, there are at least two limitations with current studies: 1) Well-labeled lung adenocarcinoma (LA, a subtype of LC) data are rare, leading to limited samples for training CNNs; 2) The conventional CNNs ignore positional information by pooling operations, whereas the positional information is of great importance in clinical diagnosis for LA. Here, we propose the "LA-Net" to address these issues by the following steps. First, we consider a transfer learning with pre-trained model based on the lung nodule (LN) classification, where the training data are richer, to assist the LA diagnosis. In addition, self-attention mechanisms are introduced to properly extract features from source dataset (LN) and to refine combined features from source and target sets for the LA classification. Moreover, we augment the CNN by another self-attention mechanism on the content and positional information. Our model has achieved 83.82% accuracy and 90.65% area under the receiver operating curve (AUC) on the LA classification task with 725 subjects, and outperforms the state-of-the-art methods. Our study supports the potential future clinical application of our method on LA diagnosis, and also suggests the importance of including domain knowledge in the design of neural networks.