Binary classification of glomeruli using convolutional neural networks
Abstract
Glomeruli are tiny clusters of blood vessels found in the kidneys which filter bodily waste. However, various diseases such as diabetes and obesity are risk factors for glomerulosclerosis, the scarring of glomeruli. This impedes the filtration process and leads to more advanced renal diseases and even kidney failure. Diagnosis currently involves a process of biopsies and classification of numerous glomeruli whole slide images. The classification of these images are time-consuming and prone to human error. Thus, there is an opportunity to utilize image processing and artificial intelligence to aid in the automation of the process, thereby aiding nephrologists to shorten diagnostic time. To this end, current literature explores the use of image classification using neural networks in the identification of glomerulosclerosis. However, existing studies make use of single convolutional neural network architectures. To improve the results, ensemble learning can be used. Ensemble learning is a method that aggregates multiple CNN architectures to reduce bias and to improve overall accuracy. Previous works demonstrated that using ensemble learning can increase accuracy in the classification of skin lesions, showing promise to its application in other fields such as glomerulosclerosis. As such, this paper uses an ensemble of CNNs to classify glomerulosclerosis in a dataset of 2,340 whole slide images of glomeruli as either "normal"or "sclerosed."Results are further optimized through the testing of different weight distributions using the weighted averaging ensemble aggregation method. Ultimately, we show that the use of ensemble learning in the binary classification of glomeruli can increase the accuracy of architectures compared to single neural networks.