DAGAN-based Gait Features Augmentation for Ankle Instability Detection.

Abstract

We developed a computer-aided program for the diagnosis of ankle ligament injury. This study used the Heidelberg Foot Measurement Method (HFMM) to obtain the specific gait kinematic characteristics. To address the issue of insufficient available medical samples, an Attention Dual Generative Adversarial Network (Attention-DualGAN) is proposed to augment the gait feature set. We used a Convolutional Long Short-Term Memory (ConvLSTM) Network-based and Time Convolution Network (TCN)-based detection model to observe the impact of the extended dataset. The results show that the accuracy of the ConvLSTM model after data enhancement improves by 2.4%, and the accuracy of the TCN model after data enhancement improves by 7.47%. Through the visualization of Pearson correlation coefficient, histogram, and scatter plot, it is proved that Attention-DualGAN produces high-quality gait kinematics characteristics. This study shows that Attention-DualGAN generates synthetic gait features with high correlation to real data, and the performance of the detection model in the diagnosis of ankle ligament injury can be improved by adding appropriate synthetic data.