John Kalkhof - PhD Data Science
Date: - Location: Eurecom
Abstract: Medical image segmentation relies heavily on large-scale deep learning models, such as UNet-based architectures, which limits their use in resource-constrained environments. In contrast, Neural Cellular Automata (NCAs) offer a lightweight and scalable alternative. Our MED-NCA model with only 13k parameters outperforms UNet-based methods, which are 500 times larger, on various MRI, CT and X-ray segmentation tasks. MED-NCA's unique one-cell architecture ensures inherent translational, scale and shape invariance and enables the use of variance as a built-in quality control mechanism. Moreover, its minimal computational demands make it possible to train directly on smartphones. By integrating MED-NCA into federated learning systems, we lower the participation threshold and democratise access to machine learning. Bio: John Kalkhof, a PhD candidate at TU Darmstadt and part of the MEC-Lab under Prof. Anirban Mukhopadhyay's guidance, is committed to advancing affordable healthcare AI to expand access to care. Initially exploring feature disentanglement during his Master's, his current research focuses on lightweight and robust segmentation methods through Neural Cellular Automata. His work was recognized at IPMI 2023 with the Francois Erbsmann Prize and with the MICCAI 2023 Young Scientist Award.
