Robust Classification of Urinary Sediment with Limited Data Through Semi-Supervised Ensemble Models

This paper proposes ensemble-based approaches to improve classification performance with limited labeled data in urine sediment images. These methods achieve significant advancements over conventional approaches. In experiments, for the case with 102 labeled data points, the proposed techniques achieved F1 scores of 0.884 and 0.901, substantially outperforming state-of-the-art baselines in both precision and recall. Classification accuracy for rare classes, such as Cryst and Cast, also improved considerably. These results emphasize the effectiveness of ensembling in reducing overfitting and variance in data-scarce settings, particularly enhancing stability and accuracy for rare particle classification. This approach holds potential for fields like medical imaging, where data is often limited, yet high classification performance is essential.