NCA-Morph: Medical Image Registration with Neural Cellular Automata

TL;DR

NC-Morph is presented, an innovative approach that seamlessly blends DL with a bio-inspired communication and networking approach, enabled by Neural Cellular Automata (NCAs), and distinguishes itself as a lightweight architecture with significantly fewer parameters.

Abstract

Medical image registration is a critical process that aligns various patient scans, facilitating tasks like diagnosis, surgical planning, and tracking. Traditional optimization based methods are slow, prompting the use of Deep Learning (DL) techniques, such as VoxelMorph and Transformer-based strategies, for faster results. However, these DL methods often impose significant resource demands. In response to these challenges, we present NCA-Morph, an innovative approach that seamlessly blends DL with a bio-inspired communication and networking approach, enabled by Neural Cellular Automata (NCAs). NCA-Morph not only harnesses the power of DL for efficient image registration but also builds a network of local communications between cells and respective voxels over time, mimicking the interaction observed in living systems. In our extensive experiments, we subject NCA-Morph to evaluations across three distinct 3D registration tasks, encompassing Brain, Prostate and Hippocampus images from both healthy and diseased patients. The results showcase NCA-Morph's ability to achieve state-of-the-art performance. Notably, NCA-Morph distinguishes itself as a lightweight architecture with significantly fewer parameters; 60% and 99.7% less than VoxelMorph and TransMorph. This characteristic positions NCA-Morph as an ideal solution for resource-constrained medical applications, such as primary care settings and operating rooms.

Figures (6)

• Figure 1: (a) Traditional optimization techniques like SyN focus on local displacements of voxels while optimizing a set of parameters during inference beg2005computingthirion1998image. (b) Traditional DL techniques leverage global information while estimating a one-shot deformation field during inference based on displacement rules learned during training balakrishnan2019voxelmorphchen2022transmorphkim2021cyclemorph. (c) NCA-Morph learns local displacement rules in a DL-fashion while regularizing the deformation field during inference based on local interactions and learned rules.
• Figure 2: NCA-Morph architecture for registration: The input is a concatenation $\odot$ of the fixed and moving image accompanied by 30 empty channels. Upon reducing the image size by a quarter, the initial NCA forecasts the overall flow of cells. Subsequently, this rough flow is combined with the detailed images, and a second NCA completes the deformation. While the second NCA solely observes patches during training, it operates at full resolution during inference. The resulting deformation field can be applied to warp both the moving image and its segmentation.
• Figure 3: NCA-Morph compared to VoxelMorph, TransMorph, ViTVNet and NICE-Trans in terms of the number of parameters and Dice performance based on the OASIS registration task. NCA-Morph uses 60% and 99.7% fewer parameters than VoxelMorph and TransMorph.
• Figure 4: Stability assessment over 10 predictions on different test samples from the OASIS registration task. Left: variance for the flow, right: variance for the corresponding segmentation masks -- using NCA-Morph$_{7\times7}^{10}$.
• Figure 5: Inference times of our NCA-Morph$_{7\times7}^{10}$ architecture across different input sizes on a Raspberry Pi.