Classification of lymphoma subtypes in PET/CT images based on a bidirectional feature fusion method

Lymphoma is a common malignancy that endangers human life and health, and accurate identification of lymphoma from PET/CT images has great value in clinical treatment. Efficient and accurate discrimination of lymphomas is an important and challenging research task. However, deep networks may lose important features such as texture structure of the target while acquiring rich image information and even misclassify small-scale targets. Traditional machine learning methods rely heavily on manually designed features and require the design of reasonable and effective feature combinations. To this end, we propose a new bidirectional feature fusion method with a lymphoma subtype classification model for PET/CT images. Firstly, deep learning latent features and machine learning explicit features of PET/CT images are extracted based on convolutional neural networks and prior knowledge, where the explicit features include distribution features and radiomics features. In the latent features extraction stage, we propose a new feature channel compression method based on squeeze-and-excitation normalization. Then, the latent features and explicit features are effectively fused based on the proposed bidirectional feature selection method. Finally, a classifier is constructed by introducing deep learning and machine learning methods for lymphoma classification. To validate the effectiveness of the model, we designed multiple sets of comparison experiments and ablation experiments to classify lymphoma subtypes, including non-lymphoma, Hodgkin's lymphoma, diffuse large B-cell lymphoma and other non-Hodgkin's lymphoma on the lymphoma dataset. The accuracy and recall of the classification reached 0.831 and 0.819, respectively. To validate the generalization of the model, we set experiments on the lung cancer PET/CT dataset, and our model improved the accuracy of classification by 0.045 compared with the resnet18 network. The experiments show that the proposed method in this paper has better classification on lymphoma dataset and can be applied to other tumors.