Dynamic Topological Routing in 3D Medical Image Segmentation: The AAM-V1 Architecture and Topological U-Net

This paper presents the AAM-V1 Topological U-Net, an advanced conditional computation framework for volumetric medical image segmentation. The architecture integrates the Medical Coherence Residual Gate (MCRG) with adaptive semantic quantization and dynamic early-exit routing, guided by the Tensor Divergence Index (TDI). TDI
serves as an entropy-driven metric to allocate computational depth and bit-precision
according to local anatomical complexity. The proposed approach achieves substantial
reductions in VRAM consumption (4–6×) and inference latency (up to 10–15× in sparse
volumes) while preserving micro-contrast in pathological regions. A machine-readable
configuration payload ensures reproducibility and deployment compatibility.
Keywords: medical image segmentation, conditional computation, dynamic neural
networks, semantic quantization, topological routing, MCRG, AAM-V1                                                   AAM-V1_ARTSYBASHEV_UA_KHARKIV_AIANALYSIS