Improving Foot Deformity Detection in Children: Synthetic Plantar Pressure Data Generation with GMM and GAN

The identification of leg deformities in children is pivotal for ensuring timely and effective medical intervention. Plantar pressure is the pressure field that acts between the foot and the support surface during everyday locomotor activities. Prior works on automatic identification of foot deformities based on plantar pressure made use of machine learning approaches. Of late, deep learning techniques have been providing state of the art performance in classification tasks. However, these algorithms require a sufficiently large labelled training data for good performance. This paper proposes two different methods for generating plantar pressure data synthetically. The first approach makes use of Gaussian Mixture Models(GMM) for generating the plantar pressure data for three distinct foot types: flat, normal, and pes cavus. The second approach makes use of one dimensional Generative Adversarial Network (GAN) for synthetic data production. The synthetically generated plantar pressure data is augmented into an existing database and subsequently classified using various machine learning approaches and one-dimensional Convolutional Neural Networks (CNN). Experiments conducted on the publicly available plantar pressure dataset show an improvement in accuracy on the incorporation of the synthetic plantar pressure data.