Classification of Ciliary Motion with 3D Convolutional Neural Networks

Abstract

Cilia are hair-like structures found on nearly every cell in the body. Motile cilia in the nasal and bronchial passages work to propel mucus and foreign matter from the respiratory system. Abnormal ciliary motion results in a wide spectrum of disorders. Currently, defective ciliary motion is diagnosed by visual examination of nasal biopsies for signs of dyskinesia. However, these examinations are highly subjective, lacking a "common language" of standardized quantitative metrics to describe the observed motion. Previous work relied on autoregressive representations of optical flow features that were fed into a support vector machine classifier. However, this approach was limited by two critical constraints: the strong parametric assumptions imposed by autoregressive models on the underlying structure of the data, and the requirement for experts to manually select regions of interest to be analyzed in the videos. Deep learning suffers from neither of these limitations. Feature maps are automatically generated from relevant portions of videos. Furthermore, the feed-forward architecture of convolutional neural networks do not impose the strong parametric assumptions of autoregressive models, permitting complex behaviors to be captured. Here, we demonstrate initial results in applying 3D deep convolutional networks to binary classification of ciliary motion. While our preliminary results lag behind those of previous work, there is considerable promise in deep learning for ciliary motion analysis and room for further experimental design.