Real-time detection transformer nodule: an improved real-time detection transformer algorithm for lung nodule detection in computed tomography images

Detecting lung nodules through computed tomography (CT) scans plays an important role in early prevention, clinical diagnosis, and monitoring of lung cancer. However, the small, irregular shape and low resolution of lung nodules, together with multiscale challenges, often hinder accurate detection, especially for small nodules. To overcome these obstacles, we propose an improved detection algorithm, a real-time detection transformer (RT-DETR) nodule, specifically for lung nodules in CT images. The proposed RT-DETR framework includes several improvements aimed at increasing the detection accuracy. First, a RFAFasterNet is designed by replacing the backbone network with FasterNet network and incorporating the receptive field attention convolution (RFAConv) module. This integration improves the network’s capacity to capture both local and distant features, guided by the receptive field attention mechanism while effectively minimizing redundant computations. In addition, a small object detection layer (P2) is added to solve the problem of low-resolution small nodule detection. Finally, a hierarchical scale–based path aggregation network is proposed to further refine the detection of different nodule sizes by integrating deep-level and shallow-level features. Comprehensive experiments on the LUNA16 dataset demonstrate that the competition performance metric reaches 88.3%, which is 2.7% higher than the baseline model. In addition, the experimental results also show that the model has strong generalization in different datasets.