Uncertainty Analysis Based Attention Network for Lung Nodule Segmentation from CT Images

In recent years, with the development of computer science and technology, computer aided diagnosis (CAD) systems have played important role in clinical practice. Using deep learning technology, existing CAD systems can predict whether CT images contain lung nodule automatically. However, for the precise segmentation of lung nodule and nodule boundary, further diagnosis by doctors is required. The current widely used segmentation networks still have segmentation uncertainty in the lung nodules boundary, which will interfere the accuracy of segmentation results. To solve this problem, this paper propose a UAA-UNet (Uncertainty Analysis Based Attention UNet) based on the uncertainty analysis of edge regions. The network structure is divided into two stages. In the first stage, the initial segmentation map of the lung nodule is obtained, and the second stage focuses on the uncertainty region of the initial segmentation map. By learning the features of the uncertainty region, the uncertainty is reduced and the segmentation accuracy is improved. The second stage includes two modules, the uncertainty attention module and the uncertainty elimination module. In uncertainty attention module, the entropy map of the initial segmentation map of the lung nodule is input into the network as attention information to improve the network's ability to understand uncertainty. In uncertainty elimination module, by using EWCE (entropy map weighted cross entropy loss function), the entropy map of the prediction result is fed back to the network as a weight factor to further improve the network's learning ability of the uncertain region. We selected lung nodule slices from 1012 patients in the Lung Image Database Consortium (LIDC) to validate the feasibility and effectiveness of the proposed method. The experiment result shows that, in the lung nodule segmentation task, by leveraging uncertainty analysis, the network achieves significant improvements over the baseline network.