Combining Multi-Scale Feature Fusion and Attention Mechanism to Achieve Pulmonary Nodule Classification

Lung cancer is the deadliest form of cancer, with most cases originating from small malignant nodules. However, the early symptoms of malignant nodules are not obvious, which can lead to misdiagnosis and delay the optimal treatment time. Therefore, differentiating between benign and malignant lung nodules is crucial for the early diagnosis and treatment of lung cancer. The advent of convolutional neural networks has made it possible to achieve rapid classification of pulmonary nodules. Most classification networks use a convolution kernel size of 3, but due to the limited receptive field, the classification of benign and malignant nodules remains inaccurate. To address this, we use larger convolution kernels to increase the receptive field, capture more detailed feature information, and enhance the ability of model to represent complex patterns and details. Furthermore, we fuse features from multiple scales to mitigate the impact of scale differences. We also employ an attention mechanism to obtain global information, highlight important areas, and suppress irrelevant parts. We use the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset to evaluate our model. Compared to other classification methods, our approach achieved superior performance, with an accuracy of 0.9855 and an F1score of 0.9901.