A Novel Automatic Lung Nodule Classification Scheme using Fusion Ghost Convolution and Hybrid Normalization in Chest CTs

Objective: To address the low efficiency of diagnosing pulmonary nodules using computed tomography (CT) images and the difficulty in obtaining the key signs of malignant pulmonary nodules, a ghost convolution residual network incorporating hybrid normalization (GCHN-net) is proposed. Methods: Firstly, a three-dimensional ghost convolution with a small kernel is embedded in the GCHN-net. Secondly, we designed a hybrid normalized-activation module (TMNAM) that can handle the rich and complex features of lung nodules in both the deep and shallow layers of the network, and incorporating two different normalization methods. This allows the network to comprehensively learn the intricate relationships underlying the intrinsic features of lung nodules and enhances its capacity to classify the properties of unknown nodules. Additionally, to enhance the accuracy and detail of the category activation map, GradCAM++ is integrated into the third layer of the GCHN-net. This integration enables the visualization of specific regions within three-dimensional lung nodules that the model focuses on during its predictions. Results: The accuracy of the GCHN-net on the Lung Nodule Analysis 16 (LUNA16) dataset was 90.22%, with an F1-score of 88.31% and a G-mean of 90.48%. Conclusion: Compared with existing methods, the proposed method can greatly improve the classification of pulmonary nodules and can effectively assist doctors in diagnosing patients with pulmonary nodules.