Bio-Measurements Estimation and Support in Knee Recovery through Machine Learning

  • 2018-07-19 16:24:22
  • João Bernardino, Luís Filipe Teixeira, Hugo Sereno Ferreira
  • 6

Abstract

Knee injuries are frequent, varied and often require the patient to undergointensive rehabilitation for several months. Treatment protocols usuallycontemplate some recurrent measurements in order to assess progress, such asgoniometry. The need for specific equipment or the complexity and duration ofthese tasks cause them to often be neglected. A novel deep learning basedsolution is presented, supported by the generation of a synthetic imagedataset. A 3D human-body model was used for this purpose, simulating arecovering patient. For each image, the coordinates of three key points wereregistered: the centers of the thigh, the knee and the lower leg. These valuesare sufficient to estimate the flexion angle. Convolutional neural networkswere then trained for predicting these six coordinates. Transfer learning wasused with the VGG16 and InceptionV3 models pre-trained on the ImageNet dataset,being an additional custom model trained from scratch. All models were testedwith different combinations of data augmentation techniques applied on thetraining sets. InceptionV3 achieved the best overall results, producingconsiderably good predictions even on real unedited pictures.

 

Quick Read (beta)

loading the full paper ...