SM-RNet: A Scale-Aware-Based Multiattention-Guided Reverse Network for Pulmonary Nodules Segmentation

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; pp. 1 - 14
• Main Authors: Tang, Tiequn, Zhang, Rongfu, Lin, Kailin, Li, Feng, Xia, Xunpeng
• Format: Journal Article
• Language: English
• Published: New York IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Attention; Cancer; Computed tomography; Datasets; Feature extraction; Image segmentation; Intermediate multiscale features (MFs); Lung; Lung cancer; lung nodule segmentation; Metric space; Nodules; Qualitative analysis; reverse erasure (RE); scale aware; scale-guided spatial attention (SSA); Segmentation; Task analysis; Tumors
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2023.3315365

Lung cancer is a leading cause of cancer death, but its mortality continues to decline substantially, benefiting from early screening and/or treatment of lung nodules. Therefore, the accurate and robust segmentation of pulmonary nodules from lung computed tomography (CT) is an essential task. However, automated solution is still challenging due to the high variability, blurred edges, and complex background of pulmonary nodules in CT scans. In this article, we propose a novel scale-aware-based multiattention-guided reverse network (SM-RNet) for pulmonary nodules segmentation, which makes full use of the complementary information from different attention dimensions and scales and progressively obtains the complete object by reverse erasure (RE) manner, consisting of two stages. Stage I performs weighted cross-scale fusion for features with different scale attentions to extract the intermediate multiscale features (MFs) with channel and scale awareness, and stage II adaptively selects features from the most appropriate scale for the target and sequentially mines the details required for the segmentation task by a new scale-guided spatial attention (SSA) block and a new RE block. In addition, the model is equipped with two-stage deep supervision. We evaluate the segmentation accuracy of our method by calculating the dice similarity coefficient (DI), Jaccard index (JI), Hausdorff distance (HD), recall, specificity (SPE), and precision (PRE) on an internal institution dataset from the Fudan University Shanghai Cancer Center (FUSCC) and a public dataset LUng Nodule Analysis 2016 (LUNA16). Our proposed SM-RNet achieves the Dice scores of 89.290% ± 0.040% and 86.496% ± 0.076% and an HD of 5.532 ± 0.186 and 6.131 ± 0.288 for FUSCC and LUNA, respectively. SM-RNet significantly outperforms other state-of-the-art (SOTA) networks. In addition, the visualization results of qualitative analysis show that the network has strong robustness.