1. A siamese network with adaptive gated feature fusion for individual knee OA features grades prediction
- Author
-
Kang Wang, Dongxing Xie, Yong Dou, Xin Niu, and Tuo Yang
- Subjects
Knee Joint ,Computer science ,Science ,Article ,Convolution ,Severity assessment ,Imaging, Three-Dimensional ,Osteoarthritis ,Bone imaging ,Humans ,Representation (mathematics) ,Feature fusion ,Multidisciplinary ,Artificial neural network ,business.industry ,Musculoskeletal abnormalities ,Pattern recognition ,Osteoarthritis, Knee ,Radiography ,Task (computing) ,Databases as Topic ,Feature (computer vision) ,Metric (mathematics) ,Medicine ,Neural Networks, Computer ,Artificial intelligence ,business ,Algorithms - Abstract
Grading individual knee osteoarthritis (OA) features is a fine-grained knee OA severity assessment. Existing methods ignore following problems: (1) more accurately located knee joints benefit subsequent grades prediction; (2) they do not consider knee joints’ symmetry and semantic information, which help to improve grades prediction performance. To this end, we propose a SE-ResNext50-32x4d-based Siamese network with adaptive gated feature fusion method to simultaneously assess eight tasks. In our method, two cascaded small convolution neural networks are designed to locate more accurate knee joints. Detected knee joints are further cropped and split into left and right patches via their symmetry, which are fed into SE-ResNext50-32x4d-based Siamese network with shared weights, extracting more detailed knee features. The adaptive gated feature fusion method is used to capture richer semantic information for better feature representation here. Meanwhile, knee OA/non-knee OA classification task is added, which helps extract richer features. We specially introduce a new evaluation metric (top±1 accuracy) aiming to measure model performance with ambiguous data labels. Our model is evaluated on two public datasets: OAI and MOST datasets, achieving the state-of-the-art results comparing to competing approaches. It has the potential to be a tool to assist clinical decision making.
- Published
- 2021