Automated Bacteriuria Detection and Grading using Deep Learning

Urine culture is a standard method for diagnosing Urinary Tract Infections (UTIs) by identifying bacteria and pus cells, but it is often time-consuming and may produce negative results. To expedite diagnosis, clinicians frequently examine urine smears under a microscope to detect bacteria before conducting the culture test. This manual approach is labor-intensive, time-consuming, and requires specialized expertise. Additionally, bacteriuria grading is essential to assess the severity of infection. Our study aims to address these challenges by automating the segmentation and grading of bacteria from gram-stained microscopic urine smear images using deep learning approaches. Although deep learning has been explored for urine smear analysis, it's potential for accurate bacterial segmentation and grading remains under explored. Deep learning models require large annotated datasets, but acquiring comprehensive data covering all grades of bacteriuria is challenging.We address these limitations in two stages. First, we propose a morphology-guided, bacteriuria-grade-aware label synthesis framework that covers all morphological classes of bacteria. Second, we propose a novel conditional GAN architecture that uses these synthesized labels to generate clinically realistic urine smear images.We used this synthetic dataset to train the proposed system. Our system demonstrated a 2% higher Fscore compared to Baseline, a 14% improvement in bacteriuria grading accuracy, and a 30% reduction in Root Mean Square Error (RMSE) for bacterial counting compared to state-of-the-art (SOTA) deep learning methods. These findings are statistically significant, indicating that our proposed approach could significantly enhance the accuracy and efficiency of UTI diagnosis, leading to improved patient outcomes.