Semi-supervised learning for hierarchical cell detection and segmentation
Machine learning is making a significant impact in many areas of computer vision including biomedical image analysis. However, state-of-the-art approaches use supervised learning that requires very large amounts of labeled data to fully constrain complex models such as convolutional neural networks. In this talk, we will discuss our recent efforts in semi-supervised learning for cell detection and segmentation. We build hierarchical models of images, i.e. trees, and use natural properties of these models to formulate an unsupervised loss function for learning. This unsupervised loss is combined with a supervised loss term in a semi-supervised setting. We experiment with our method for cell detection in light microscopy images and for neuron segmentation in electron microscopy images and show that the proposed unsupervised loss function allows us to obtain highly accurate results even when only a few labeled samples are available for learning.