Cilia Segmentation in Medical Videos with Fourier Convolutional Neural Network

The study of cilia motion analysis is crucial because cilia are found on almost all vertebrate cells. Abnormal function of cilia can manifest as a variety of symptoms. However, manually segmenting the cilia area is time consuming as well as difficult even to a professional. The distinct hair-like structure of cilia and the periodical beating motion of cilia makes it possible to bring Fourier transform to neural network to largely help with separating the cilia area with other areas. Fourier convolution operator and pooling operator are proposed. The octave convolution blocks are introduced to reduce the memory issue faced by the Fourier convolution operator. Two models are proposed in this paper to deal with the problem of cilia segmentation. The first model is an entirely unsupervised approach in the way of basic structure of W-Net network with Fourier convolution blocks. The second model is a partial reconstruction model with Fourier convolution blocks for reconstruction of videos with cilia area only. Several problems are demonstrated in the result and these should be explained in future work.