MED-NCA: Bio-inspired medical image segmentation.

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Computationally intensive models like U-Net and Transformers create a healthcare technology divide, limiting access in low-resource regions.
• This disparity particularly affects medical diagnostics and treatment planning in low- and middle-income countries, primary care, and conflict zones.

Purpose of the Study:

• Introduce MED-NCA, a Neural Cellular Automata (NCA) based segmentation model designed for efficiency and robustness in resource-constrained environments.
• Extend the validation of MED-NCA across diverse medical imaging modalities and anatomies.
• Enhance model interpretability and robustness testing with a novel visualization tool, NCA-VIS.

Main Methods:

• Developed MED-NCA, a low-parameter Neural Cellular Automata (NCA) model for medical image segmentation.
• Validated MED-NCA across eight distinct anatomies and imaging types (MRI, CT, X-ray, Ultrasound, 2D, 3D).
• Introduced NCA-VIS for visualizing MED-NCA inference and testing robustness against artifacts.

Main Results:

• MED-NCA demonstrated comparable performance to significantly larger U-Net models across various medical imaging tasks.
• The model's low parameter count enables efficient operation on minimal hardware, such as a Raspberry Pi or smartphone.
• NCA-VIS provided insights into the model's inference process and robustness.

Conclusions:

• MED-NCA offers a transformative, efficient, and accessible solution for medical image analysis, particularly in resource-limited settings.
• The technology has broad applicability across diverse anatomies and imaging modalities.
• MED-NCA and NCA-VIS contribute to advancing global healthcare equity by democratizing advanced diagnostic capabilities.