Attention-Guided 3D Residual Learning for Pulmonary Embolism Segmentation on CT Pulmonary Angiography.
Authors
Affiliations (4)
Affiliations (4)
- Department of Chest Diseases, Faculty of Medicine, Dicle University, Diyarbakır, Türkiye.
- Department of Electrical and Electronics Engineering, Faculty of Engineering, Bingöl University, Bingöl, Türkiye. [email protected].
- Department of Computer Engineering, Faculty of Engineering, Dicle University, Diyarbakır, Türkiye.
- Department of Computer Engineering, Faculty of Engineering and Architecture, Bingöl University, Bingöl, Türkiye.
Abstract
Pulmonary embolism is a potentially life-threatening condition that requires timely and accurate diagnosis. Computed tomography pulmonary angiography (CTPA) is the standard imaging modality for pulmonary embolism assessment; however, manual delineation of embolic regions is time-consuming and challenging because emboli are often small, irregular, and embedded within complex contrast-enhanced vascular structures. This study aimed to develop and evaluate a patch-based 3D deep learning model for automated pulmonary embolism segmentation on CTPA. A retrospective CTPA dataset was collected from Dicle University Hospital with institutional ethical approval. After image-mask quality control, 272 subjects were included and split using a subject-based strategy into training, validation, and test subsets. Lesion-centered positive patches and negative-control patches were extracted to construct a 3D segmentation dataset consisting of 816 patches. We propose PEARL-Net3D (pulmonary embolism attention-guided residual learning network in 3D), a compact residual 3D encoder-decoder architecture incorporating convolutional block attention modules and attention-gated skip fusion. The model was compared with 3D U-Net, 3D VNet, and 3D ResUNet using Dice/F1, IoU, precision, recall, HD95, and ASSD. Ablation experiments and paired Wilcoxon signed-rank tests were also performed. PEARL-Net3D achieved the highest test-set Dice/F1 and IoU among all evaluated models, with mean values of 0.8287 ± 0.2782 and 0.7809 ± 0.3175, respectively. Compared with the strongest baseline, 3D ResUNet, PEARL-Net3D improved Dice/F1 by 0.0098 and IoU by 0.0098, while also achieving the lowest ASSD of 3.0840 ± 3.3025. Statistical analysis showed significant improvements over 3D ResUNet in Dice/F1 (p = 0.0133), IoU (p = 0.0137), and recall (p = 0.0080). Ablation analysis demonstrated that the combined use of CBAM and attention-gated skip fusion yielded the best validation performance, with a Dice/F1 of 0.8486 and IoU of 0.8026. PEARL-Net3D provides effective patch-based 3D segmentation of pulmonary embolism on CTPA and outperforms standard 3D segmentation baselines in overlap-based metrics. The proposed architecture improves embolus voxel recovery and boundary agreement by combining residual volumetric feature extraction with attention-based feature recalibration and gated skip fusion.