Multi-Head Attention for Pulmonary Embolism Detection in CT Angiography

Pulmonary embolism (PE), often caused by deep vein thrombosis, is a critical condition with high morbidity and mortality. Timely diagnosis through CT Pulmonary Angiography (CTPA), the gold standard imaging modality, is vital but hampered by the time-intensive and variable nature of manual interpretation. Deep learning models offer potential for automated PE detection but struggle with smaller, peripheral emboli, leading to underdiagnosis. This study addresses these challenges by employing advanced deep learning architectures: nnUNET, Swin Transformer, and our proposed MHA-UNET. These models focus on comprehensive detection across all pulmonary artery branches, emphasizing subtle, peripheral emboli. The MHA-UNET integrates multi-head attention mechanisms to enhance focus on critical regions, while the Swin Transformer leverages hierarchical attention for multi-scale feature analysis. Results demonstrate that MHA-UNET outperforms other models, achieving superior sensitivity, specificity, and robustness in detecting small and peripheral emboli. This work establishes a framework for automated, precise emboli detection, aiming to improve clinical outcomes and radiologist efficiency.