HierDETR: Hierarchical Multi-Head Attention-based Pulmonary Nodule Detection in Medical Imaging

Abstract

Lung cancer is one of the leading causes of morbidity and mortality worldwide, and early detection is crucial for improving patient survival rates. Medical imaging, particularly chest CT scans, plays a vital role in early lung cancer screening and pulmonary nodule detection. However, the detection of pulmonary nodules is challenging due to the variability in size, shape, density, and location of nodules, as well as their complex background, which often includes normal or benign structures such as blood vessels and bronchial walls. To address these challenges, this paper presents a novel pulmonary nodule detection framework based on Hierarchical Multi-Head Attention (HMHA) and Query-Key Cache Updating (QKCU) mechanisms. The proposed HierDETR model improves upon existing methods by reducing redundancy in attention heads and effectively utilizing multi-scale and hierarchical context information. Experimental results on the publicly available LUNA16 dataset demonstrate that the HierDETR model outperforms mainstream methods, achieving significant improvements in detection metrics such as F1 Score, Average Precision, and Average Recall. This work provides a promising approach for enhancing the robustness and accuracy of pulmonary nodule detection, with potential applications in clinical practice for early lung cancer diagnosis.