HCGN: Hierarchical Convolution and Graph Network for Predicting Knee Osteoarthritis

EasyChair Preprint 15330

Abstract

Knee osteoarthritis (KOA) is a common joint disease that severely affects the normal lives of patients. Typically, in clinical practice, the severity of KOA is evaluated by observing X-ray images of the knee joint. This approach is highly dependent on the subjective experience of the doctor and may vary among doctors. In this study, we propose a deep convolutional neural network (CNN) model that integrates structural information processing to predict KOA severity automatically based on the Kellgren-Lawrence (KL) grading system. Specifically, (1) The knee joint regions of the original X-ray images are segmented using automatic detection for subsequent model predictions; (2) We employed popular pre-trained deep CNN models to perform feature extraction, obtain their multi-scale feature maps, and construct their corresponding graph representations; (3) A graph attention network (GAT) was designed as a fine-tuning module to build a KOA prediction model. In our experiments, we tested various pretrained models combined with a GAT fine-tuning module to evaluate their performance on the Osteoarthritis Initiative (OAI) dataset. The results show that our proposed method significantly improves the predictive performance in multiple aspects compared to the original model. In addition, our proposed method has good decision interpretability.(https://github.com/ddw2AIGROUP2CQUPT/HCGN)

Keyphrases: Knee Osteoarthritis, feature space, graph structure, image classification

@booklet{EasyChair:15330,
  author    = {Xionghui Yang and Pengju Tang and Kai Zou and Dawei Dai},
  title     = {HCGN: Hierarchical Convolution and Graph Network for Predicting Knee Osteoarthritis},
  howpublished = {EasyChair Preprint 15330},
  year      = {EasyChair, 2024}}