Objective: Pulmonary embolism (PE) is a life-threatening condition requiring rapid and accurate diagnosis. This study proposes a deep learning-based approach for automated PE segmentation, focusing on both pixel-level accuracy and clinical applicability. The objective is to enhance segmentation performance by incorporating Squeeze-and-Attention (SA) modules within the network architecture and to evaluate its generalizability on multi-center datasets.

Materials and methods: A neural network architecture combining convolutional layers, Long Short-Term Memory (LSTM) units, and SA blocks was developed. The model was trained and evaluated using two publicly available PE datasets and two private datasets. Performance was assessed using five-fold cross-validation, hold-out testing, and ablation studies. Evaluation metrics are computed at pixel, slice, and lesion levels, both per-patient and per-slice.

Results: The results demonstrated a Dice similarity coefficient of 69.87% at the pixel level and 98.79% sensitivity at the lesion-level on the public dataset, indicating competitive segmentation accuracy. The incorporation of SA blocks significantly improved performance, increasing the Dice score by 7.58% while reducing both false positive and false negative rates. The model showed good generalizability across different imaging centers and scanner types.

Conclusion: The proposed model demonstrates acceptable accuracy in PE segmentation and robustness across diverse datasets. Its design and evaluation framework support its potential utility in clinical settings. However, accurate delineation of small emboli remains challenging, and further investigation is required to improve segmentation performance in such cases. Future work will focus on real-time deployment and integration into diagnostic workflows.