关键词: 肺结节检测, 注意力机制, 特征金字塔网络, 假阳性抑制, 卷积神经网络

Abstract:

Aiming at the problems of low sensitivity and high false positive rate caused by various shapes and difficulty in detecting pulmonary nodules by the Computer-Aided Detection （CAD） system of pulmonary nodules， a pulmonary nodule detection algorithm based on attention feature pyramid networks was proposed. In the first stage， a more compact Dual Path Network （DPN） was used as the backbone network， and a Feature Pyramid Network （FPN） was combined for multi-scale prediction to obtain feature information at different levels. At the same time， the Global Attention Mechanism （GAM） was embedded to refine the semantic features to be emphasized in learning and improve the sensitivity of the algorithm. In the second stage， a false positive reduction network was proposed to obtain the final classification prediction results. In the training stage， the focal loss function and various data augmentation techniques were used to deal with the data imbalance problem. Experimental results on the public dataset LUNA16 （LUng Nodule Analysis 2016） show that the Competitive Performance Metric （CPM） of the algorithm only with the first stage reaches 0.908， and after adding the false positive reduction network， the CPM of the algorithm reaches 0.933， which is 1.1 percentage points higher than that of the classic algorithm — Convolutional Neural Network （CNN） based on Maximum Intensity Projection （MIP）. And ablation experimental results show that the dual path network， FPN， and GAM are effective in improving the detection sensitivity. The above proves that the proposed two-stage detection algorithm can obtain multi-scale nodule information， improve the sensitivity of pulmonary nodule detection， and reduce the false positive rate.

Key words: pulmonary nodule detection, attention mechanism, Feature Pyramid Network (FPN), false positive reduction, Convolutional Neural Network (CNN)