Implantable Adaptive Cells in U-Net Skip Connections: Low-Compute Retrofit for Medical Segmentation

We present a retrofit neural architecture search (NAS) framework that upgrades pre-trained medical image segmentation networks by searching only the skip-connection modules while keeping the encoder-decoder backbone frozen. The method insertsImplantable Adaptive Cells(IACs), small differentiableNAS- found plug-ins discovered with a PC-DARTS-inspired procedure tailored to the skip-only setting. Unlike standard NAS, which produces a new model trained end-to-end from scratch, our approach preserves existing weights and refines a deployed U-shaped architecture in place. Across four MRI/CT benchmarks (ACDC, AMOS, BraTS, KiTS) and six 2D U-Net backbones, IACs consistently improve patient-level Dice over the reference U-Net while adding only ≈ 7.6% parameters, ≈ 14% FLOPs and ≈ 19% latency on average. Compared to strong hand-crafted skip refinements (Attention Gates, Residual Attention blocks, U-Net++), IACs achieve larger or comparable Dice gains at similar computational cost. When plugged into the 2D nnU-Net pipeline, IACs further raise foreground Dice by +0.02 to +0.37 without data augmentation and by +0.05 to +0.15 with augmentation, with all improvements statistically significant at the patient level. To stabilize differentiable search, we modify edge normalization via a capped rescaling that promotes sparse, unambiguous edge selections and eases the continuous-to-discrete transition. Search is performed once per pre-trained backbone on a single GPU and takes hours rather than days; ablations show that the budget can be reduced by nearly a factor of five with minimal loss in accuracy. Overall, IACs provide a practical, low-compute path to upgrading deployed U-shaped models when data or compute are limited. Code is available at https://gitlab.com/emil-benedykciuk/u-net-darts-tensorflow.