VoxLogicA2 speeds medical image analysis with logic and AI fusion
VoxLogicA2: A Spatial Model Checker for Neuro-symbolic Medical Imaging (extended version)
Logic in Computer Science
Summary
Analyzing medical images like brain scans can be very complex and slow. The authors present VoxLogicA2, a software tool that combines logical rules with neural networks to quickly and accurately analyze 2D and 3D images. This tool lets users write programs that handle both learning from data and applying logical spatial checks in one place. The authors show their tool is faster and more efficient than previous methods, making medical image analysis easier to specify and run.
What this means in practice
- •For medical imaging teams: Combine AI and logical analysis to accelerate processing and verification of medical scans for diagnosis and treatment planning.
- •For computer vision developers: Integrate spatial reasoning with neural network models in one program to build more explainable image processing workflows.
Authors
Gina Belmonte, Laura Bussi, Vincenzo Ciancia, Mieke Massink
Abstract
We present the tool VoxLogicA2, a spatial model checker aimed at Computer Vision, with applications in high-resolution 2D and 3D medical imaging. The new version of the tool combines logical primitives, non-logical primitives, and training and prediction primitives of state-of-the-art neural networks, freely interleaved with the symbolic tool functionality. The primary application field is neuro-symbolic medical image analysis. VoxLogicA2 is a fast, parallelising spatial model checker with a declarative domain-specific language, capable of describing both the learning procedures and the spatial logic procedures of an analysis in one program. We validate our approach by verifying its improved efficiency, and by using the new constructs of the language to describe case studies of previous work, which were based on Python, with a specification that uses VoxLogicA2 only