Lung Nodule Classification Based on SE-ResNet152 and Stratified Sampling

Lung nodules, an early indication of lung cancer, are crucial for its treatment. Existing studies primarily focus on improving model structures, neglecting the issue of data imbalance in lung nodule classification. In this work, we propose a multiple-stage stratified sampling (MS-SS) to address the issue of data imbalance. This approach aims to achieve data balance while preserving the original data distribution structure to the maximum extent. Additionally, we introduce the SE-ResNet152 model combined with transfer learning to handle lung nodule classification, enabling feature recalibration through the SE module. To evaluate the proposed method, experiments are conducted on the Luna16 dataset. The results demonstrate a remarkable F1-Score of 96.358% on the test set, confirming the effectiveness of our approach in accurately classifying lung nodules.