نمونه کار Noval Method Medical Image Classification

• Hybrid Medical Image Classification & Retrieval Engineered a hybrid CNN-ViT architecture integrating CoAtNet backbones with parallel Side-ViT branches via gated cross-attention fusion; utilized Proxy Anchor Loss for discriminative embedding learning and achieved SOTA efficiency using composite PEFT strategies. • Model Architecture: CoAtNet (MBConv + Self-Attention), Vision Transformer (ViT), Hierarchical Hybrid CNN-ViT, Feature Pyramid Networks (FPN), Cross-Attention Fusion, Gated Attention Modules, Squeeze-and-Excitation (SE) Blocks. • PEFT & Optimization: Low-Rank Adaptation (LoRA), Additive Side-Tuning, Fine-Grained Prompt Tuning, Stochastic Depth, DropBlock, Gradient Checkpointing, Automatic Mixed Precision (AMP).