Symbolic and hybrid AI for brain tissue segmentation using spatial model checking

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Neuroimaging

Background:

• Manual segmentation of 3D medical images, particularly brain tumors, is time-consuming and lacks reproducibility.
• Accurate and explainable segmentation methods are crucial for neuroimaging and radiotherapy applications.
• Current segmentation techniques often struggle to balance accuracy, efficiency, and interpretability.

Purpose of the Study:

• To develop a novel symbolic approach for automated brain segmentation and lesion delineation.
• To create a hybrid AI method combining symbolic logic with machine learning for enhanced segmentation.
• To provide an accurate, explainable, and efficient alternative to manual segmentation in neuroimaging.

Main Methods:

• A novel symbolic approach based on spatial model checking, utilizing closure spaces and spatial logics.
• Development of a declarative logic language (ImgQL) and a spatial model checker (VoxLogicA).
• Integration of the symbolic method with Machine Learning techniques to form a hybrid AI approach.

Main Results:

• The symbolic approach demonstrated accurate segmentation on multiple public 3D magnetic resonance (MR) image datasets (BraTS Challenges, BrainWeb).
• The hybrid AI method achieved state-of-the-art segmentation accuracy and computational efficiency.
• The proposed methods provide explainable segmentation results, a key advantage over existing techniques.

Conclusions:

• The novel symbolic and hybrid AI approaches offer accurate, efficient, and explainable solutions for 3D brain segmentation.
• This work addresses the challenge of automated tumor delineation, improving upon manual methods.
• The findings have significant implications for neuroimaging, radiotherapy, and clinical decision-making.