Currently, pulmonary nodules detection work mostly focus on recognition and diagnosis of solid nodules. However, ground glass nodules have higher probability of malignancy, posing greater identification challenges and thus greater value for detection. To achieve rapid and accurate detection of ground glass nodules. This article proposed an algorithm based on RT-DETR model with the following enhancement: 1) optimize the backbone network with FCGE blocks to increase the detection accuracy of small-sized and blurred edge nodules
2) replace the AIFI module with HiLo-AIFI module to reduce redundant computation and improve the detection accuracy of pure ground glass pulmonary nodules and mixed ground glass pulmonary nodules
3) replace the DGAK module with CCFF module to address the issue of capturing complex features and recognition of irregularly shaped ground glass nodules. To obtain a more lightweight model, modules are designed for smaller number of parameters and higher computational efficiency. Model are tested on mixed dataset composed of LIDC-IDRI data and clinical data from cooperating hospitals. Compared to the baseline model, it shows an average precision improvement (mAP50/mAP50:95) of 2.1% and 1.7%, with a reduction parameters by 5.2 million. On a specialized dataset containing both pure and mixed ground glass nodules, our model outperformed the baseline model in all evaluation metrics. In general, the model proposed in this paper achieves improvement on lightweightness and detection accuracy. However, the model exhibits poor noise resistance and robustness, suggesting optimization in future work.