Abstract:A precise and efficient pulmonary nodule detection model is crucial for the accurate early screening of lung cancer. However, current algorithms suffer from problems such as low detection accuracy and insufficient robustness when dealing with small-scale pulmonary nodules and those with special structures. Based on the YOLO11 algorithm and by integrating dual-domain features, this study proposes an improved DFT-YOLO11 (dual-domain (spatial-frequency) discrete Fourier transform-enhanced YOLO11) model for precise identification and localization of pulmonary nodules. The DFT-YOLO11 model innovatively introduces the frequency-domain-spatial-domain feature alignment attention mechanism module and the FC3k2 module that integrates the frequency-domain branch, providing richer semantic feature information for pulmonary nodule detection from the perspective of the frequency domain. Meanwhile, the proposed adaptive threshold loss function balances the influence of nodules at different scales during network training. The experimental results show that the DFT-YOLO11 model achieves a mAP0.5 of 94.45% on the LUNA16 dataset, surpassing existing pulmonary nodule detection models. On the Lung-PET-CT-Dx dataset, it reaches a mAP0.5 of 97.29%, demonstrating the capability of DFT-YOLO11 for multi-class cancerous nodule detection. Moreover, the morphological structural indexes of pulmonary nodules calculated from DFT-YOLO11 detection results yield the lowest mean squared error, providing reliable support for clinical applications.