A Novel Deep Learning Framework Based Mask-Guided Attention Mechanism for Distant Metastasis Prediction of Lung Cancer

Abstract

Distant metastasis (DM) is the leading cause of death in advanced lung cancer, which is diagnosed by positron emission tomography (PET) scanning. Compared with the expensive price and nocuous contrast medium in PET, using computed tomography (CT) for DM diagnosis is more economical and convenient in clinical practice. However, most existing methods only analyze tumor regions to extract local features for DM prediction, which neglects the rich whole-lung information. To alleviate this problem, we propose a novel deep learning framework based mask-guided attention mechanism called Mask-Guided Two-stream Attention network (MGTA) for DM prediction, including a 3D pseudo-siamese feature pyramid network (PSFPN) to learn both global features in the whole lung and local features in tumor; and a deep cascaded attention module (DCAM) for further feature fusion. The proposed MGTA enjoys several merits. First, to the best of our knowledge, this is the first work to simultaneously mine tumor and whole-lung information through a mask-guided mechanism for DM prediction. Second, the proposed deep cascade attention module can effectively leverage the complementary multi-level features in PSFPN to enhance the feature recognition capacity of small tumors. Extensive experiments on a large-scale DM dataset including 2814 lung cancer patients show that the MGTA achieves good performance with area under the curve (AUC) of 0.822; sensitivity of 0.753; and specificity of 0.743, which outperforms state-of-the-art lung cancer diagnosis methods and the commonly used tumor-based methods. Furthermore, the MGTA shows large improvement especially when the training data size is small.