An Optimized Data and Model Centric Approach for Multi-Class Automated Urine Sediment Classification

Abstract

Automated urine sediment analyzers play a crucial role in diagnosing urinary tract infections, offering real-time data analysis and expediting patient diagnosis. This paper introduces a novel hybrid approach combining data-centric and model-centric techniques for automated urine sediment analysis. The proposed methodology addresses challenges such as morphological similarities among particle classes, uneven particle distribution, and intra/inter-class variations. A modified version of convolutional neural network (CNN), specifically the Visual Geometry Group (VGG-19) model, incorporating transfer learning, along with data augmentation is proposed for automated urine sediment classification with 98% accuracy and impressive inference time of 61ms per image. The proposed approach outperforms existing methods, especially in handling diverse sediment categories, demonstrating its potential for practical applications in medical diagnostics. We proposed the integration of a data-centric approach for improved labeling reliability and a model-centric approach for fine-tuning of the deep learning model, showcasing promising results in recognizing 12 distinct urine sediment classes. This study also emphasizes the importance of collaboration with medical professionals in refining the model's performance and handling challenges related to data acquisition and class imbalance. The proposed approach provides a significant advancement in automating and enhancing urine sediment analysis processes.