Nick Lemke

This review highlights Neural Cellular Automata as efficient, robust, and lightweight alternatives to traditional neural networks for medical imaging, examining their architectures, applications, limitations, and future research opportunities.

StyleGANCA is a lightweight GAN that generates high-quality, class-informative medical images with far fewer parameters than existing generative models, achieving state-of-the-art results on PathMNIST with just 617k parameters.

SG-NCA enables accurate, privacy-preserving surgical scene understanding on fanless edge devices with performance comparable to larger models while using 55× fewer parameters.

TheODen enables secure, firewall-compatible federated learning for histopathology, demonstrating robust cancer segmentation across multiple German hospitals without sharing sensitive data or opening client-side ports.

UNEG, a multi-model benchmark for continual learning in medical image segmentation, outperforms existing methods by maintaining separate networks for each training stage and using reconstruction error to select the appropriate model during inference.

FedNCA enables secure, low-cost federated medical image segmentation on resource-constrained edge devices, helping overcome computing, connectivity, and privacy barriers to AI healthcare adoption in LMICs.

OctreeNCA enables fast, globally consistent segmentation of extremely large medical images and videos at once, using about 90% less GPU memory than a UNet.

Neural cellular automata can deliver high-quality medical image segmentation with dramatically fewer parameters, enabling training and inference on low-cost devices like smartphones and microcontrollers.

A distribution-aware replay strategy for medical image segmentation mitigates forgetting through feature auto-encoding while detecting model failure from out-of-distribution instances, addressing privacy concerns and unexpected distribution shifts.