Training a Supervised Cilia Segmentation Model from Self-Supervision

Cilia are organelles found on the surface of some cells in the human body that sweep rhythmically to transport substances. Dysfunctional cilia are indicative of diseases that can disrupt organs such as the lungs and kidneys. Understanding cilia behavior is essential in diagnosing and treating such diseases. But, the tasks of automatically analysing cilia are often a labor and time-intensive since there is a lack of automated segmentation. In this work we overcome this bottleneck by developing a robust, self-supervised framework exploiting the visual similarity of normal and dysfunctional cilia. This framework generates pseudolabels from optical flow motion vectors, which serve as training data for a semi-supervised neural network. Our approach eliminates the need for manual annotations, enabling accurate and efficient segmentation of both motile and immotile cilia.