Hybrid dynamic routing architecture for high-efficiency Tuberculosis classification. By cascading a lightweight CNN with a fine-tuned Vision Transformer (ViT), the system optimizes the latency-accuracy trade-off in real-time.

An entropy-based router directs easy samples to the CNN and only forwards low-confidence samples to the compute-heavy ViT. This approach minimizes computational overhead without compromising diagnostic precision.

Key Performance Metrics:

• 2.2x Latency Reduction: Optimized mean inference time from 140ms (ViT-only) to 62ms.

• Efficient Offloading: 66% of samples are resolved via the high-speed CNN path.

• High Fidelity: Maintained >99% accuracy on the held-out test set.