A multi-head YOLOv12 with self-supervised pretraining for urinary sediment particle detection

Abstract

Automatic and reliable urine sediment analysis is essential for timely diagnosis and management of renal and urinary disorders. However, manual methods are time-consuming, subjective, and limited by operator abilities. In this study, we propose a novel deep learning method based on a multi-head YOLOv12 architecture combined with self-supervised pretraining and advanced inference through Slicing Aided Hyper Inference (SAHI) to effectively address these challenges. Unlike prior methods that employed a single detection head, our architecture features six specialized and independent detection heads: Cells, Casts, Crystals, Microorganisms/Yeast, Artifact, and Others, enabling simultaneous and fine-grained classification of the full spectrum of urine sediment particles, including all relevant subclasses. To facilitate robust training, we created a large-scale dataset (OpenUrine) encompassing 790 labeled images with over 31,285 bounding boxes across 39 categories, and 5640 unlabeled images for self-supervised learning. Evaluated on this complex 39-class dataset, our model achieved a precision of 76.59% and a mean Average Precision (mAP) of 64.15%, demonstrating competitive performance in detection accuracy, especially of small and low-contrast objects.