A Novel Approach to Pulmonary Embolism Segmentation: Increasing an Attention-Based U-Net

Abstract

Pulmonary embolism (PE) is a life-threatening condition, often leading to late diagnoses. Diagnostic tools like Computed Tomography Pulmonary Angiography (CTPA) rely on radiologist skills, resulting in variable sensitivity and specificity. This study aims to leverage deep learning, specifically a convolutional neural network with U-net architecture enhanced by Squeeze-and-Attention and Long Short-Term Memory (LSTM) blocks, to improve the segmentation of emboli in CTPA images. Utilizing two datasets, CAD-PE (91 cases, 89 with PE) and FUMPE (35 cases, 33 with PE), the research assesses how increasing the number of network layers (57, 67, and 103) affects segmentation performance. The results demonstrated that the slice-wise sensitivity improved from 76.73 pm 21.94 with a 57 layer architecture to 80.36 pm 21.42 with a 67-layer architecture, indicating better pulmonary embolism detection (with a significant difference due to paired T-test P-value of less than 0.05). In addition, the patient-wise AUC slightly increases from 81.68 pm 10.94 (57 layers) to 85.09 pm 10.69 (67 layers) with a Kruskal-Wallis P-value of 0. 0 1 8 9, which indicates a significant difference between the networks' performance. However, no significant improvement was observed with the 103-layer model, highlighting the potential for overfitting. Results from this study demonstrate the potential of deep learning algorithms in enhancing the accurate diagnosis of pulmonary embolism. Finally, the neural network's performance in segmenting pulmonary embolisms from CT images demonstrates promising directions with particular specificity and overall AUC strengths. © 2025 IEEE.