A Novel Attention-based Neural Network for Automated Lung Lesion Delineation from 4DCT Images

Radiotherapy plays an important role in the treatment of non-small cell lung cancer. Accurate segmentation of the gross target volume (GTV) is critical for successful radiotherapy delivery. Deep learning techniques are able to offer fast and accurate segmentation, which could serve as decision support for specialists and save time compared to manual delineation. Regarding the large the volume of the lung compared to the lung lesions, achieving lesion segmentation with very high accuracy is challenging due to the large input field. In this light, dedicated networks with attention modules are needed for automated and accurate lung lesion delineation. This work used a 3D-Unet equipped with a novel attention module to segment the gross tumor volume from 4D CT images presenting with non-small cell lung cancer. A total of 730 clinical 4DCT lung images were included in this study, 657 for training and 73 images for test. The proposed framework consists of two major components: attention and segmentation modules, wherein the attention module would estimate approximate location of the lung lesion through a coarse segmentation of the input image with high true-positive rate to ensure the inclusion of the entire lesion (using a dedicated loss function). The outcome of the attention module as well as the key feature maps from the decoder part are fed into the segmentation module. Given the restricted field provided by the attention module, the segmentation module would perform accurate/precise delineation of the lung lesion.The DSC, JSC and Hassdorf distance indices were used to evaluate our proposed method. Dice score increased from 0.75 to 0.83. Jac-card and Hausdorff distance indices also improved from 0.62 to 0.73, and from 4.29mm to 3.73mm, respectively. The proposed network, exhibited superior performance to the network without attention model.