Semi-Supervised Representation Learning for Infants Biliary Atresia Screening Using Deep CNN-Based Variational Autoencoder

Abstract

The challenge of convolutional networks (CNNs) for medical imaging analysis is to train the network model with limited well labeled dataset. Since a variational autoencoder (VAE) is able to learn the probability distribution on data for describing an observation in terms of its latent attributes by unsupervised manner, it has emerged as one of the most popular unsupervised learning technology in computer vision applications. In this paper, we presented a semi-supervised representation learning approach for screening infant’s biliary atresia using convolutional variational autoencoder (CVAE). Firstly, we leveraged a smartphone’s camera for infant’s stool images collection. Secondly, a pre-trained deep convolutional variation autoencoder was used to train the feature extractor for the infant’s stool image-features extraction, and finally we fine-tuned the last classification layers for identifying acholic from normal stools. We compared our screening approach with “tradition stool color card” method; the results demonstrated CVAE model has a higher accuracy rate 92.16%. Universal screening biliary atresia by semi-supervised representation learning may be a valuable technology to help parents identify acholic stools in the perinatal period, which may ultimately lead to improved native liver survival probabilities.