GLOMNET: A Hover Deep Learning Model for Glomerulus Instance Segmentation

Glomeruli are essential kidney structures for blood filtration. Their damage can impact the filtering capability of the kidney, leading to its failure. Hence, glomerulus detection and evaluation are crucial for kidney disease diagnosis. Most deep learning methods for glomerulus segmentation focus on semantic segmentation and do not allow instances’ separation. In this paper, we present GlomNet, a network for glomerulus instance segmentation. Our network is composed of one EfficientNet encoder and two decoders: one for binary output prediction and the other for the prediction of Horizontal and Vertical distances of object pixels to their centers of mass (HoVer Maps). During post-processing, these HoVer Maps are used to determine object boundaries and separate glomerular instances. Our method was trained and tested on 176 images from two clinical centers. Our network outperformed two state-of-the-art methods with a Dice score of 0.79 and a Panoptic Quality of 0.61.