Semantic Segmentation of Glomeruli in Human Kidney Tissue Images

Abstract

For reliable disease diagnosis in renal pathology, accurate glomerular microscopic medical image segmentation is important. In this study, we work on a pixel-level labeled glomerular microscopic medical image segmentation dataset, the HuBMAP kidney dataset, and improve a novel pipeline for implementing automatic segmen- tation of glomerular microscopic medical images. In our thesis work, we have tried using many variations of segmentation models, encoders, feature extractors and explored their potentials for semantic segmentation of glomeruli. In our proposed approach, we used the network architecture which gives the most promising re- sult on the dataset, consisting of LinkNet with E cientNet as modi ed encoder block, pretrained on ImageNet. Here the Compound Scaling provides better per- formance without compromising the e ciency. This pipeline outperformed other models that we have experimented with and allowed better performance than previous non deep learning based methodologies of glomerular identi cation.