Lesion Feature Extraction and Classification Optimization Method Using Dynamic Fusion of Global Attention and Local Attention

In tumor diagnosis, due to subtle differences in the imaging appearance of different diseases, accurately classifying lesions based on solely imaging data proves challenging. Existing machine learning and deep learning methods face limitations due to the small sample size of medical datasets and the intricate nature of disease image manifestations. This paper proposes a novel lesion classification method to fully explore distinctions among confused lesion features associated with different diseases. The proposed method comprises three key steps: Firstly, a lesion feature calculation method using dynamic fusion of global attention and local attention is proposed. The weight of global attention and local attention is dynamically allocated, and the global and local features are fused by dynamic weight. Secondly, feature dimension reduction is realized to improve the effect of distinguishable features using sparse autoencoder and polynomial constraint loss function. Finally, to improve the performance of classification, the monarch butterfly optimization algorithm based on adaptive neighborhood search radius method is used to optimize the parameters of multi-kernel support vector machine. The private PET/CT image classification dataset of lymphoma and Still’s disease was used to validate our results. The experimental results demonstrate that the method's efficacy in lymphoma and Still’ disease classification tasks, achieving an accuracy (ACC) of 82.8% and an area under the curve (AUC) of 87.1%, respectively.