Automatic Urine Sediment Detection and Classification Based on YoloV8

Abstract

The identification of urine sediment in human urine samples through microscopic images is a critical part of in vitro testing. Currently, automatic urine sediment analyzers are used by doctors to supplement manual examinations. However, the conventional technique of artificial feature extraction used by most analyzers can be labor-intensive and subjectively dependent on the professional’s prior knowledge. To overcome these limitations, this work employs YoloV8, a recent version of the Yolo algorithm, to accurately detect and categorize urine particles. In addition, a data-centric strategy has been introduced to address difficulties with missing data, incorrect labeling, and class imbalance. This strategy aims to improve labeling reliability and remove noisy data points. Experimental findings on the dataset show that YOLOv8 has a greater detection accuracy than existing state-of-the-art techniques for detecting eleven different categories of urine sediments. The approach presented in this work outperforms other techniques, yielding a mean average precision (mAP) of 91%. Furthermore, the average detection time of the model is 0.6 microseconds.