Disentangled Sequential Autoencoder with Local Consistency for Infectious Keratitis Diagnosis

Yuxuan Si, Zhengqing Fang, Kun Kuang, Zhengxing Huang, Yu-Feng Yao, Fei Wu

Abstract

Infectious keratitis is a major cause of visual impairment and a common blinding eye disease. Deep learning based prior researches mainly regard infectious keratitis diagnosis as a classification task on the slit-lamp images of single-visit. However, in real clinical applications, it is critical to analyze the lesion evolution characteristics represented by time-varying features over multiple-visits. To bridge this gap, in this paper, we focus on the problem with sequential clinical images of patients, and propose a novel disentangled sequential auto-encoder (DSLC-VAE) algorithm to separate the time-varying pathological features from the time-invariant ones for infectious keratitis diagnosis. Specifically, a inference model is exploited to generate time series of the shape and appearance of corneal lesions to represent keratitis progression, which are combined with location-related features to identify keratitis pathogen. Moreover, we construct a local consistent regularizer with a self-supervised task to enhance the consistency of the time-varying features across different infectious keratitis. Extensive experiments on real world dataset demonstrate superiority of our DSLC-VAE on both representation disentanglement and diagnosis accuracy.
BibTeX
@inproceedings{si2022disentangled,
  author = {Yuxuan Si and Zhengqing Fang and Kun Kuang and Zhengxing Huang and Yu-Feng Yao and Fei Wu},
  title = {{Disentangled Sequential Autoencoder with Local Consistency for Infectious Keratitis Diagnosis}},
  booktitle = {ICIP},
  year = {2022},
  doi = {10.1109/icip46576.2022.9897646},
}

Loading related content...

Figure