DAG-Net: Dual-Branch Attention-Guided Network for Multi-Scale Information Fusion in Lung Nodule Segmentation

The development of deep learning has played an increasingly crucial role in assisting medical diagnoses. Lung cancer, as a major disease threatening human health, benefits significantly from the use of auxiliary medical systems to assist in segmenting pulmonary nodules. This approach effectively enhances both the accuracy and speed of diagnosis for physicians, thereby reducing the risk of patient mortality. However, pulmonary nodules are characterized by irregular shapes and a wide range of diameter variations. They often reside amidst blood vessels and various tissue structures, posing significant challenges in designing an automated system for lung nodule segmentation. To address this, we have developed a three-dimensional dual-branch attention-guided network (DAG-Net) for multi-scale information fusion, aimed at segmenting lung nodules of various types and sizes. First, a dual-branch encoding structure is employed to provide the network with prior knowledge about nodule texture information, which aids the network in better identifying different types of lung nodules. Next, we designed a structure to extract global information, which enhances the network's ability to localize lung nodules of different sizes by fusing information from multiple resolutions. Following that, we fused multi-scale information in a parallel structure and used attention mechanisms to guide the network in suppressing the influence of non-nodule regions. Finally, we employed an attention-based structure to guide the network in achieving more accurate segmentation by progressively using high-level semantic information at each layer. Our proposed network achieved a DSC value of 85.6% on the LUNA16 dataset, outperforming state-of-the-art methods, demonstrating the effectiveness of the network.