Microscopic urinary particle detection by different YOLOv5 models with evolutionary genetic algorithm based hyperparameter optimization

Abstract

The diagnosis of kidney disease often involves analysing urine sediment particles. Traditionally, urinalysis was performed manually by collecting urine samples and using a centrifuge, which was prone to manual errors and relied on labour-intensive processes. Automated urine sediment microscopy, based on machine learning models, requires segmentation and feature extraction, which can hinder model performance due to intrinsic characteristics of microscopic images. Deep learning models based on convolutional neural networks (CNNs) often rely on a large number of manually annotated data, making the system computationally complex. This study propose an advanced deep learning model based on YOLOv5, which offers faster performance and requires comparatively less data. The proposed model used five variants of the YOLOv5 model (YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x) to detect six categories of urine particles (erythrocyte, leukocyte, crystals, cast, mycete, epithelial cells) from microscopic urine sediment images. The dataset involved 5376 images of urine sediments with 6 particles. There are 30 sets of hyperparamreteres are employed in the YOLOv5 model. To optimize the hyperparameters and fine-tune with the urine sediment dataset and for training each model, the system employed a genetic algorithm (GA) based on evolutionary principles named as Evolutionary Genetic Algorithm (EGA). Among the six categories of detected particles mycete achieved maximum performance with a mAP of 97.6 % and crystals achieved minimum performance with a mAP of 81.7 % with YOLOv5x model compared to other particles. To optimize the hyperparameters for training each model, the system employed a genetic algorithm (GA) based on evolutionary principles named as Evolutionary Genetic Algorithm (EGA). Among all the models, YOLOv5l and YOLOv5x performed the best. YOLOv5l achieved a mean average precision (mAP) of 85.8 % while YOLOv5x achieved a mAP of 85.4 % at an IoU threshold of 0.5. The detection speed per image was 23.4 ms for YOLOv5l and 28.4 ms for YOLOv5x. The proposed method developed a faster and better automated microscopic model using advanced deep learning techniques to detect urinary particles from microscopic urine sediment images for kidney disease identification. The method demonstrated strong performance in urinalysis.