Medical image classification remains a major challenge due to medical image complexity (dispersed pathological features, inter-class similarity, and intra-class variability). Existing hierarchical multi-scale feature fusion frameworks (HiFuse) have achieved good performance by integrating global and local features via parallel branches, but there is still a need to capture richer multi-level information for better feature representation. Inspired by this, we propose HiFuse++, which enhances classification accuracy by introducing two new feature fusion branches and a novel fusion module. Specifically, we add two branches to the original HiFuse architecture to extract complementary features at different semantic scales for richer contextual information. We also design a lightweight fusion module integrated into the output layer and hierarchical feature fusion (PPFM) module, enabling hierarchical feature refinement, adaptive interaction between original and new branches, and suppression of redundant information. Experiments on three medical image datasets (ISIC2018, COVID-19-CT, and Tetralogy of Fallot, TOF) validate HiFuse++’s superiority: compared with baseline HiFuse, it achieves 4% higher accuracy on ISIC2018 and 5.1% on TOF while maintaining computational efficiency. The results show that the multi-branch architecture and enhanced fusion strategy significantly improve feature discriminability, offering a robust solution for medical image analysis.