Dual-channel grouped cross-dimension attention V-Net for pulmonary nodule segmentation.

Background: Accurate segmentation of pulmonary nodules is crucial for the diagnosis and treatment of early-stage lung cancer as it can aid clinicians in formulating effective treatment plans, increasing the chance of early detection and treatment, and reducing mortality. However, pulmonary nodules are similar to surrounding tissues, and the location, size, and quantity of nodules in different patients are unpredictable, posing challenges to accurate segmentation. This study aimed to develop a new deep learning network based on V-Net to address the deficiencies in pulmonary nodule segmentation tasks. Methods: This work proposes a dual-channel grouped cross-dimension attention V-Net (DGCA V-Net) model for computed tomography (CT) pulmonary nodule segmentation. In downsampling, the model uses the global grouped coordinate attention (GGCA) module to comprehensively capture multidimensional global information and reduce the loss of feature information. In the bottom decoding path, the grouped split attention (GSA) module is adopted to effectively reduce the loss of detailed information. A dual-input guided feature aggregation (DGA) module is introduced between the encoding and decoding paths to effectively alleviate the impact of inaccurate localization learning on the detailed segmentation of pulmonary nodules. Results: The proposed model was trained and evaluated with the Lung Nodule Analysis 2016 (LUNA16) pulmonary nodule public dataset. The segmentation performance was statistically analyzed with four indicators: the Dice similarity coefficient (DSC), intersection over union (IoU), precision, and recall. The proposed DGCA V-Net model significantly outperformed the baseline model V-Net in all metrics: the DSC increased by 4.72% to 0.7921, the IoU increased by 6.92% to 0.6662, the precision increased by 2.90% to 0.8102, and the recall increased by 4.80% to 0.7993. Conclusions: Based on experimental data, the strategy presented in this study significantly improves lung nodule segmentation accuracy. The ablation tests also confirm the proposed module's strong segmentation and generalization abilities. This segmentation model is expected to be applied to other medical segmentations. Our solution is open-source and online (https://github.com/freshmancodes/Pulmonary-Nodule-Segmentation).