A classification study of pulmonary nodule CT images based on supervised contrast learning

With the increasing amount of pulmonary odule CT images and the difficulty of labeling categories, traditional deep learning methods that rely on labeled samples may not be able to learn effectively. To address these challenges, this paper proposes a 3D convolutional network for pulmonary nodule feature extraction using contrastive learning. The proposed method consists of two phases: discriminative feature learning and classification. In the discriminative feature learning phase, a supervised contrast learning loss is used to reduce the distance between nodules in the same class while increasing the distance between nodules in different classes, thus improving intra-class diversity and inter-class similarity. In the subsequent classification phase, pre-training weights are fine-tuned to better fit the LUNA16 dataset. The proposed network achieves an accuracy of 86.96, an F1-score of 81.28, and a mean G-mean of 83.86 on the LUNA16 dataset. Compared with existing supervised learning methods, introducing contrastive learning allows for the full use of labeled data and improves the classification accuracy of lung nodules with the same amount of training data.