Unsupervised learning for labeling global glomerulosclerosis

Current deep learning models for classifying glomeruli in nephropathology are trained almost exclusively in a supervised manner, requiring expert-labeled images. Very little is known about the potential for unsupervised learning to overcome this bottleneck. To address this open question in a proof-of-concept, the project focused on the most fundamental classification task: globally sclerosed versus non-globally sclerosed glomeruli. The performance of clustering between the two classes was extensively studied across a variety of labeled datasets with diverse compositions and histological stains, and across the feature embeddings produced by 34 different pre-trained CNN models. As demonstrated by the study, clustering of globally and non-globally sclerosed glomeruli is generally highly feasible, yielding accuracies of over 95% in most datasets. Further work will be required to expand these experiments towards the clustering of additional glomerular lesion categories. We are convinced that these efforts (i) will open up opportunities for semi-automatic labeling approaches, thus alleviating the need for labor-intensive manual labeling, and (ii) illustrate that glomerular classification models can potentially be trained even in the absence of expert-derived class labels.