MICCAI 2026 Tutorial • Thursday, October 1, 2026 • 8:00-12:00

Learning to Self-Organize: Neural Cellular Automata for Medical Imaging

Organizers: John Kalkhof, Caroline Essert, Marco Lorenzi, Carsten Marr, Anirban Mukhopadhyay, Camila González, Marawan Elbatel, Daniel Lang, Ario Sadafi, Maximilian Buser, Michael Deutges, Jonas Mehtali, Henry Krumb, Mirko Konstantin, Nick Lemke

  • Duration~4h (incl breaks)
  • FormatLecture ↔ Hands-on (~50/50)
  • DeliveryGoogle Colab
  • RequirementsLaptop

Abstract

Medical image analysis increasingly relies on large neural networks. While powerful, these models often lack robustness across diverse data, are difficult to interpret, and are inefficient for large-scale or edge computing applications. Neural Cellular Automata (NCA) are an emerging architectural paradigm: they learn local update rules that are iteratively applied so that complex anatomical or morphological patterns emerge through self-organization. This formulation enables models to reach performance comparable to much larger architectures while using two to three orders of magnitude fewer parameters and providing insight into the intermediate updates. This tutorial introduces the foundations of NCAs, their training and emergent behavior, and reviews recent medical-imaging applications including segmentation, temporally consistent synthesis, robust classification, and fast surrogate modeling for therapy planning. Prepared Colab-based hands-on sessions will guide participants to train models, visualize local updates, and apply NCAs to example tasks.

View Agenda Access Materials

Why This Tutorial

Medical image analysis increasingly relies on large neural networks, which raises challenges around computational efficiency, robustness across heterogeneous data, and interpretability in clinical practice. NCA approaches these problems through local iterative updates that produce global structure via self-organization.

The tutorial is designed for participants familiar with deep learning but new to NCAs, and aims to provide both conceptual understanding and practical experience. By the end, participants should be able to train NCAs for medical imaging tasks, visualize their iterative dynamics, and evaluate trade-offs between performance, robustness, and efficiency.

Agenda

10 min08:00-08:10Welcome
20 min08:10-08:30Foundations
30 min08:30-09:00Hands-on 1
20 min09:00-09:20Applications
40 min09:20-10:00Hands-on 2
30 min10:00-10:30Break
20 min10:30-10:50Simulation
10 min10:50-11:00Tips
30 min11:00-11:30Hands-on 3
15 min11:30-11:45Outlook
15 min11:45-12:00Q&A
Duration Time Session Speaker Topic
10 min 08:00-08:10 Lecture John Kalkhof

Welcome and setup (Wi-Fi check, Colab access, goals)
20 min 08:10-08:30 Lecture John Kalkhof

Foundations of NCA (CA → NCA, local updates, emergence, architecture)
30 min 08:30-09:00 Hands-on Michael Deutges

Train a minimal NCA in Colab and visualize iterative updates
20 min 09:00-09:20 Lecture Daniel Lang

Applications in medical imaging (segmentation, registration, detection)
40 min 09:20-10:00 Hands-on Marawan Elbatel

Train classification and segmentation models; evaluate robustness
30 min 10:00-10:30 Break ☕
20 min 10:30-10:50 Lecture Caroline Essert

NCA for simulation
10 min 10:50-11:00 Lecture Jonas Mehtali

Tips for training and technical details
30 min 11:00-11:30 Hands-on Jonas Mehtali

Extend to simulation
15 min 11:30-11:45 Lecture Ario Sadafi

Interpretability and outlook
15 min 11:45-12:00 Discussion All

Final Q&A

Materials

Open the materials for each hands-on session.