Dual-channel CNN-Transformer architecture combining DenseNet-169 and Swin-Tiny for three-class tumor detection (non-affected, benign, malignant) on 224×224 RGB images. DenseNet branch feeds through frozen backbone layers DB1–DB2, progressively unfreezing DB4 at epoch 10 and DB3 at epoch 20, while Swin-Tiny unfreezes Stage 3 and Stage 2 on the same schedule. Both branches output 512-dimensional embeddings via projection layers, fused through a gated attention module that learns optimal channel weighting before classification. MEAB block in the DenseNet path applies multi-dilated convolutions, squeeze-excitation attention, and adaptive feature fusion to capture multi-scale tumor morphology. Training uses cosine annealing (50 epochs), label smoothing, balanced class sampling via stratified 80/20 split, and early stopping on validation accuracy with 10-epoch patience. Fork this diagram on Diagrams.so to customize layer unfreezing schedules, adjust fusion gate mechanisms, or adapt for different medical imaging datasets and classification tasks.