SSD-VUPV: A Novel SSD-Based Model for Automated Pulmonary Nodule Detection

urpose: In the context of increasingly scarce medical resources, the main purpose of this paper is to propose a SSD-based Variance-based Upsampling and Pyramid Voting (SSD-VUPV) model for improving the efficiency and accuracy of automated detection of lung nodules, as one of the early manifestations of lung cancer. Methods: Firstly, the proposed SSD-VUPV model adapts to the detection of small pulmonary nodules by changing the size of the feature map of the input prediction module. Secondly, the prediction frame is modified to make greater use of the shallow feature layer. In addition, a set of up-Block modules is augmented through the incorporation of asymmetric convolutions, which involves the utilization of a Feature Pyramid Network (FPN) mechanism. Finally, multi-scale and asymmetric convolutions are added to the model to further improve its detection performance. Results: Experimental results, obtained on two public datasets, show that SSD-VUPV can increase the [email protected], F1 score, and sensitivity of the baseline model (SSD) from 63.01% to 87.52%, 64.20% to 90.24%, and 63.25% to 88.74%, on the LUNA16 dataset, and from 77.6% to 83.2%, 76.5% to 83.0%, and 75.8% to 81.9%, on the ChestX-ray14 dataset, respectively. Moreover, SSD-VUPV outperforms state-of-the-art models, based on their results reported in the literature, according to all evaluation metrics used. Conclusion: By cleverly integrating a feature pyramid structure and incorporating the newly designed up_Block modules, the proposed SSD-VUPV model can combine deep semantic features with shallow detail features, thus fully leveraging the rich feature information in the medical images, which allows it to reach top detection accuracy and robustness. Moreover, the inclusion of the newly designed Visual Multi-scale Asymmetric Convolution (VMAC) modules enables the model to adapt to different scale receptive fields, which enables it to capture more varied and detailed features, deepening its understanding of the input data and significantly improving its ability to capture features of various sizes. Consequently, the proposed SSD-VUPV model exhibits improved performance in scenarios involving complex backgrounds and targets.