Prediction of Plantar Shear Stress Distribution by Conditional GAN with Attention Mechanism

Abstract

Diabetic foot ulcers (DFUs) are known to have multifactorial etiology. Among the biomechanical factors that lead to plantar ulcers, shear stresses have been either neglected or unmeasured due to challenges in complexity and equipment availability. The purpose of this study is to develop a software that predicts plantar shear stress using plantar pressure and temperature distributions. Thirty-one subjects, 8 of them at risk of developing DFUs were recruited, and plantar thermography, pressure and shear stress distributions were collected. We introduce the conditional generative adversarial networks (cGAN) for shear stress distribution prediction and propose an attention mechanism to improve the model’s accuracy. The networks can learn the mapping from pressure to shear stress distribution. The attention mechanism can merge temperature distribution into GAN without resizing or aligning it manually. We then test on our dataset with 185 groups. The predicted anteroposterior shear stress distributions give \(72.97\%\) accuracy on peak location prediction and 14.12 kPa on global root mean square error. Our initial results are promising in terms of feasibility of our approach in predicting plantar shear stresses and this approach may benefit to address the DFU risks before ulceration.