Minimal annotation training for segmentation of microscopy images
Damian J. Matuszewski, Ida‐Maria Sintorn · 2018
In many biomedical applications, successful training of Convolutional Neural Networks (CNNs) is restricted by an insufficient amount of annotated images. Although image augmentation can help training CNNs from a relatively small image set, in many applications, the objects of interest cannot be accurately delineated due to their fuzzy shape, image quality or a limitation in time, experience or knowledge of the expert performing the annotation. We propose an approach for training a CNN for segmentation of images with minimal annotation. The annotation consists of center points or lines of target objects of approximately known size. We demonstrate this approach in the application of Rift Valley virus segmentation in a challenging transmission electron microscopy image dataset. Our method achieves a Dice score of 0.900 and intersection over union of 0.831. Using the suggested minimal annotation training is particularly useful for applications in which full object annotations are not available or feasible.