Deep Neural Nets: Pediatric Hand Radiographs


  • Matthew Chen
  • David B. Larson, MD, MBA
  • Matthew P. Lungren, MD, MHA
  • Safwan Halabi, MD
  • Curtis P. Langlotz, MD, PhD

Training Details

(4363 images)


  • Deep Residual Network
  • 50 layers
  • Adam SGD optimizer
  • Decaying learning rate
  • Pre-trained Weights
Larson, DB, Chen, MC, Lungren, MP, Halabi, S, Langlotz, CP. Performance of a Deep Neural Network Model to Assess Skeletal Maturity on Pediatric Hand Radiographs. Annual Conference on Machine Intelligence for Medical Imaging (SIIM MIMI). Society for Imaging Informatics in Medicine, Alexandria, VA, 2016

Data Pre-Processing

Training Time

  • Random rotations
  • Random crops
  • Random contrast
  • Random Left Right Flips
  • Mean subtraction

Test Time

  • Ten fixed crops
  • Mean subtraction