Gliomas have the highest mortality rate of all brain tumors. Correctly classifying the glioma risk period can help doctors make reasonable treatment plans and improve patients’ survival rates. This paper proposes a hierarchical multi-scale attention feature fusion medical image classification network (HMAC-Net), which effectively combines global features and local features. The network framework consists of three parallel layers: The global feature extraction layer, the local feature extraction layer, and the multi-scale feature fusion layer. A linear sparse attention mechanism is designed in the global feature extraction layer to reduce information redundancy. In the local feature extraction layer, a bilateral local attention mechanism is introduced to improve the extraction of relevant information between adjacent slices. In the multi-scale feature fusion layer, a channel fusion block combining convolutional attention mechanism and residual inverse multi-layer perceptron is proposed to prevent gradient disappearance and network degradation and improve feature representation capability. The double-branch iterative multi-scale classification block is used to improve the classification performance. On the brain glioma risk grading dataset, the results of the ablation experiment and comparison experiment show that the proposed HMAC-Net has the best performance in both qualitative analysis of heat maps and quantitative analysis of evaluation indicators. On the dataset of skin cancer classification, the generalization experiment results show that the proposed HMAC-Net has a good generalization effect.

Aiming at the problems of information loss and the relationship between features and target tasks in multimodal medical image segmentation, a multimodal medical image segmentation algorithm based on feature decoupling and information bottleneck theory is proposed in this paper. Based on the reversible network, the bottom-up learning method for different modal information is constructed, which enhances the features’ expression ability and the network’s learning ability. The feature fusion module is designed to balance multi-directional information flow. To retain the information relevant to the target task to the maximum extent and suppress the information irrelevant to the target task, the feature decoupling module is designed to ensure a strong correlation between the feature and the target task. A loss function based on information bottleneck theory was intended to improve information quality and remove redundant information. Based on BraTs2021, BraTs2023-MET and ANNLIB datasets, the proposed algorithm is analyzed qualitatively and quantitatively in this paper. In the quantitative experiment, the Dice coefficient of the proposed algorithm was increased by 0.110 on average compared with other methods, and the HD95 was decreased by 28.568 on average compared with other methods. In qualitative analysis, the proposed algorithm can effectively segment the incoherent region between the lesion and the lesion boundary and achieve accurate segmentation of the lesion.