Virtual Phase-Contrast Microscopy for Urinary Sediment Analysis: Deep-Learning-Based Image Translation from Brightfield Microscopy

Phase-contrast microscopy enhances the visualization of transparent or low-contrast urinary sediment elements but requires specialized optical components. In this study, we develop and evaluate a supervised U-Net-based image-to-image translation model to generate virtual phase-contrast images from brightfield urinary sediment images. A total of 1124 paired brightfield and true phase-contrast images acquired using the UriSed 3 system are classified into 898 training, 113 validation, and 113 test images. The model used a composite loss combining mean absolute error (MAE), structural similarity index measure (SSIM), and gradient preservation and was compared with a simple convolutional reference model. Performance was evaluated using root mean square error percentage (RMSE%), MAE, peak signal-to-noise ratio (PSNR), SSIM, and histogram cosine angle. Blinded expert evaluation included 45 comparison sets across nine urinary sediment categories. The proposed model outperformed the comparator in terms of RMSE% (1.3558 vs. 3.3195), MAE (0.0079 vs. 0.0276), PSNR (38.4257 vs. 29.9917 dB), SSIM (0.9573 vs. 0.9003), and histogram cosine angle (42.3470° vs. 87.0299°). Expert evaluation suggested that the generated virtual phase-contrast images were visually and clinically interpretable. These findings support the use of virtual phase-contrast generation as a supportive visualization tool for urinary sediment analysis, although external validation is required before clinical implementation.