슈퍼픽셀 분할 수준에 따른 그래프 어텐션 신경망 기반 영상 분류 성능 평가

Abstract

Intelligent Process Automation (IPA), which integrates robotic process automation with artificial intelligence, has been increasingly adopted across various industries. However, its application in medical imaging remains at an early stage. This study proposes a medical image classification system based on a Graph Attention Network (GAT) to detect pneumonia from chest X-ray images. To effectively capture local structural information, each input image is segmented into various numbers of superpixels using the Simple Linear Iterative Clustering (SLIC) algorithm. These superpixels serve as graph nodes, each represented by a feature vector. Edges are assigned using dilation to reflect neighboring relationships. The proposed GAT architecture incorporates multi-head attention, batch normalization, and dropout layers, and its performance was evaluated using stratified 5-fold cross-validation. Results showed that when the number of superpixels was set to 100, the model achieved its best performance with an accuracy of 87.7% and an F1-score of 88.2%, along with low standard deviations, indicating robust and consistent outcomes. In contrast, testing with 300 superpixels revealed a trade-off between precision and recall. These findings demonstrate that the proposed model shows potential for sensitive detection of pneumonia cases; however, excessive over-segmentation may hinder its generalizability.