Unsupervised learning for labeling global glomerulosclerosis

Labeling images for supervised learning in nephropathology is highly time-consuming and dependent on domain-expertise. Unsupervised strategies have been suggested for mitigating this bottleneck. For instance, previous work suggested that clustering/grouping of glomeruli based on image features might enable a more semi-automated labeling of morphological classes or even a completely unsupervised training. However, even for the most basic separation between globally sclerosed and non-globally sclerosed glomeruli, the performance of clustering approaches has not yet been fully elucidated. The current study sought to fill this gap by extensively evaluating the accuracy and limitations of capturing these two classes via clustering.

Clustering was investigated across 10 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.

While further work will be required to expand these experiments towards the clustering of additional glomerular lesion categories, the study clearly demonstrates that clustering might serve as a highly accurate means of pre-labeling glomeruli. Importantly, these findings strongly support clustering as a solid basis for downstream interactive labeling approaches or unsupervised learning approaches. Together, these results might greatly improve the possibilities and lookout for the establishment of clinically applicable glomerular classification models in the community. Further improvements in this area might be achieved by exploring more domain-specific feature extractors through contrastive learning or established foundation models.