Lesion feature enhancement and boundary-aware fusion for pulmonary embolism recognition on CT images.

Objective: To address the challenges of complex lesion representation, blurred vascular boundaries, and insufficient model generalization ability in pulmonary embolism CT images, an automatic recognition method that combines lesion feature enhancement and boundary-related structural cue fusion is proposed. Methods: An end to end pulmonary embolism recognition framework was constructed. On the basis of hierarchical visual encoding, a Pulmonary Embolus Feature Enhancement Module and a Vascular Boundary Aware Fusion Module were introduced to strengthen lesion discriminative representation and feature-level boundary-related structural modeling, respectively. Comparative experiments, ablation studies, visualization analysis, and external dataset validation were conducted to systematically evaluate the effectiveness of the proposed method in the binary classification setting. Results: On 523 single-center chest CT cases, the proposed method achieved an Accuracy of 0.956, a Precision of 0.961, a Recall of 0.951, and an AUC of 0.971, outperforming multiple comparison models in the internal evaluation. In external validation, the proposed method achieved Accuracy values of 0.779 and 0.672 on the FUMPE and RSNA datasets, respectively, indicating its potential cross-domain generalization ability in pulmonary embolism recognition. Conclusion: The proposed method can effectively improve the discriminative performance of binary pulmonary embolism recognition from CT images, and may provide a useful technical reference for computer-aided screening of pulmonary embolism.