Classification of glomerular spikes using Convolutional Neural Network
Yilin Chen, Ming Li, Hao Fang, Weixia Han, Dan Niu, Chen Wang · 2020
Membranous nephropathy (MN) is a common cause of adult nephrotic syndrome. In clinical, MN is diagnosed by pathological biopsy. Under the light microscope, one of the main features is that a large number of immune complex deposits can be seen on the epithelial side of the glomerular capillary loop, which is called spikelike projections. However, the spikes are very small and the diagnosis of spikes requires pathologist's experience, so it is easy to cause missed diagnosis. In this paper, we proposed a model, SPIKE-NET, for the classification of glomeruli, which combine the U-Net and the ResNet. U-Net is used as a data preprocessing step to segment the glomeruli, and ResNet is used to classify the spiked glomeruli and normal glomeruli. We established a dataset containing 1267 images of glomeruli stained with PASM. The F1 of SPIKE-NET reaches 0.9448. The higher Recall of the model means a lower missed diagnosis rate, which is of great significance in clinical diagnosis and provides a good foundation for the intelligent auxiliary diagnosis of membranous nephropathy.