PRGNN: A Position-Aware Regional Graph Neural Network for Brain PET Image Classification

PRGNN is an innovative graph neural network architecture specifically designed for brain PET image classification tasks, capturing spatial relationship features in medical images through a position-aware mechanism. This architecture divides brain images into anatomical regions as graph nodes, models the spatial relationships and metabolic pattern interactions between regions, aiming to address the limitations of traditional CNNs in modeling long-range spatial relationships.

Positron Emission Tomography (PET) is an important imaging technique in neuromedicine for detecting brain metabolic activity, which is of great value for the early diagnosis of neurodegenerative diseases. However, PET image analysis faces challenges: fine spatial correlations between different brain regions under complex three-dimensional structures, and lesions are often metabolic abnormalities in specific regions rather than global changes. Traditional CNNs excel at local texture features but have limitations in modeling long-range spatial relationships and inter-regional interactions.

The core idea of PRGNN is to divide brain images into meaningful anatomical regions as graph nodes, and model the spatial relationships and metabolic pattern interactions between regions through graph neural networks. Its key innovation lies in the position-aware mechanism: distinguishing features and their occurrence positions through spatial coordinate embedding, relative position encoding, and position-feature fusion; the regional graph construction adopts a data-driven strategy, where nodes include features such as regional metabolic intensity and texture statistics, edges are established based on anatomical proximity and functional similarity, capturing both local structural and global functional network features.

PRGNN uses an improved message-passing mechanism. When nodes aggregate neighbor features in each graph convolution layer, relative positional relationships are considered, making message passing spatially selective. The model introduces an attention mechanism to dynamically adjust connection weights, which can highlight brain region connections that contribute significantly to diagnostic decisions, having interpretability value.

PRGNN provides a new path for automated analysis of brain PET images and is expected to generate value in the following aspects: early disease screening (capturing subtle metabolic changes and identifying pathological signs before clinical symptoms); auxiliary diagnostic decision-making (attention weights provide interpretable biological basis to help doctors understand the logic of AI decisions); longitudinal monitoring (position-aware characteristics are suitable for tracking metabolic changes in specific brain regions over time, evaluating disease progression or treatment effects).

PRGNN reflects the trend of medical AI evolving from general models to domain-specific architectures. Understanding the characteristics of imaging modalities and clinical needs is no less important than expanding model scale. The idea of position-aware graph neural networks can be extended to 3D imaging tasks such as CT and MRI multi-organ segmentation. Integrating anatomical priors into models in the form of graph structures is a direction to improve interpretability and reliability. With the accumulation of multi-center datasets and the development of federated learning, such methods are expected to learn from large-scale data under privacy protection, enhancing generalization ability and clinical practicality.