Glomerulus Semantic Segmentation Using Ensemble of Deep Learning Models

Quantification and classification of tissue features such as glomerulus are important elements of the histopathologic assessment of renal tissue. To fulfill this task, glomerulus segmentation is required. In this paper, we propose a multi-stream glomerulus segmentation framework based on three signature models: FCN, Deeplabv3 and Unet. Resnet is combined with FCN and Deeplabv3 model to enhance the encoding process. Meanwhile, Unet is upgraded to use EfficientNet as the backbone for feature extraction. Each individual model will output a local decision. On top of these base learners, ensemble approaches are proposed for robust performance through prediction aggregation. Among all the ensemble methods, the Bayesian voting method performs best and achieves F-score of 91.5%