Learning to Segment Blob-like Objects by Image-Level Counting

There is a high demand for manually annotated data in many of the segmentation tasks based on neural networks. Selecting objects pixel by pixel not only takes much time, but it can also lead to inattentiveness and to inconsistencies due to changing annotators for different datasets and monotonous work. This is especially, but not exclusively, the case with sensor data such as microscopy imaging, where many blob-like objects need to be annotated. In addressing these problems, we present a weakly supervised training method that uses object counts at the image level to learn a segmentation implicitly instead of relying on a pixelwise annotation. Our method uses a given segmentation network and extends it with a counting head to enable training by counting. As part of the method, we introduce two specialized losses, contrast loss and morphological loss, which allow for a blob-like output with high contrast to be extracted from the last convolutional layer of the network before the actual counting. We show that similar high F1-scores can be achieved with weakly supervised learning methods as with strongly supervised training; in addition, we address the limitations of the presented method.