Due to the intricate dynamic coupling between molecular networks and brain regions, early diagnosis and pathological mechanism analysis of Alzheimer's disease (AD) remain highly challenging. To address this, we propose a graph convolutional neural network framework (DEG-BRIN-GCN) based on a differentially expressed gene-brain region interaction network (DEG-BRIN), aiming to enhance both diagnostic accuracy and biological interpretability in AD research. We began by systematically analyzing transcriptomic data from 19 brain regions, identifying 329 differentially expressed genes that display widespread co-expression across multiple regions. Using these findings, we constructed DEG-BRIN to model prior associations among genes, thereby revealing potential molecular connectivity patterns implicated in AD pathological progression. Leveraging this network prior, we developed an AD classification model based on graph convolutional networks. Comparative experiments demonstrate that our proposed DEG-BRIN-GCN achieves significantly better diagnostic performance than three categories of baseline models: traditional machine learning methods, Random-GCN (models based on random network topologies), and PPI-GCN. Further analysis identified key brain regions–such as the superior parietal lobule, putamen, and frontal pole–along with high-contribution genes, including VCAM1, MCTP1, HBB, and CX3CR1, which play critical roles in AD pathology. Notably, this study is the first to implement a interpretability analysis based on a “gene-region-pathway” triad, offering a novel framework for cross-scale exploration of AD pathological mechanisms. Our findings underscore the central importance of inter-regional molecular interaction networks in the accurate diagnosis of AD.
Human pose estimation (HPE) is a fundamental problem in computer vision, aiming to obtain the spatial coordinates of human joints from images or videos. Despite significant progress, traditional methods often struggle with real-time performance due to their computational complexity. In this paper, we propose a lightweight 3D human pose estimation and visualization system based on the BlazePose+SYS model, which employs an encoding-decoding architecture. Our system can detect human key points and reconstruct 3D human poses in real-time, providing high-precision and real-time solutions for various applications such as action recognition, virtual reality, and sports training. By discarding the heatmap branch during inference, our model achieves lightweight performance suitable for mobile devices and edge computing environments. Experimental results on public datasets demonstrate the effectiveness and accuracy of our system in real-time human pose estimation tasks. This work contributes to advancing the field of HPE by providing a practical and efficient solution.