Automatic segmentation of malignant pulmonary nodules on LUNA16 subsets using an attention-guided 3D U-Net framework

Objective Lung cancer is the leading cause of cancer-related mortality worldwide, and accurate segmentation of malignant pulmonary nodules is essential for quantitative lesion analysis and computer-aided lung cancer screening. Manual segmentation is time-consuming and highly dependent on radiologist expertise, highlighting the need for accurate and efficient automated segmentation techniques. Methods This study proposes an attention-guided 3D U-Net framework for volumetric segmentation of malignant pulmonary nodules based on the LUNA16 dataset. The model integrates attention gates into skip connections, adopts a hybrid Dice-BCE loss to alleviate class imbalance, and applies safe 3D data augmentation to improve segmentation performance. The network outputs a 3D binary segmentation mask with the same size as the input patch. Systematic ablation experiments and comparative evaluations were performed against nine state-of-the-art segmentation models under standardized protocols. Results On the independent test set, the proposed method achieves a Dice similarity coefficient of 0.8432, IoU of 0.7301, Precision of 0.8526, Recall of 0.8894, and AUC of 0.942, exhibiting superior overall segmentation performance compared with all competing models. Conclusion The proposed automated segmentation framework provides accurate and consistent segmentation results for malignant pulmonary nodules under the evaluated dataset. It demonstrates effective segmentation capability for small, low-contrast, and pleura-adhered malignant nodules, showing potential application value in computer-aided lung cancer screening and providing a technical basis for further clinical assistance.