A Review on Various Machine Learning Algorithms for Classification of Urine Sediments

Background and aims: Analyzing urine can reveal the presence of a variety of issues and illnessesin the human body. It is a well-known fact that patients’ urine is collected in order to diagnose avariety of illnesses, particularly diabetes, metabolic, urinary, and kidney diseases. Also, informationconcerning the presence of infection in the urine is provided by urine culture. Traditionally,sediment swabs are examined under a microscope, and the particles are counted and sorted manually.This can be a time- consuming and labor-intensive process. It also fails to meet the standardsof today’s clinics and is subject to human error and observer skill. An automated method that canassess and quantify urine sample images would be very helpful to clinicians and patients.Method: We searched Science Direct and Google Scholar for papers on machine learning studiesfor urine sediment classification published between January ۱, ۲۰۱۸, and March ۳۱, ۲۰۲۳. Wesummarized them into categories based on the following axes: year of publication, data types,sample size, algorithm used, preprocessing methods, types of urine sediments, and model performance.We selected ۳۸ articles in total for the comparison among different types of machinelearning algorithms for urine sediment classification.Results: These literatures reviewed the several machine learning algorithms used for automaticrecognition of urine particles, including erythrocytes, cylinders, leukocytes, crystals, bacteria,yeast, sperm, casts, epithelial cells, bacteria, RBC, WBC, etc. from urine sediment microscopicimages. Based on the literature review, it turned out that SVM, AdaBoost, KNN, different CNNarchitectures, DFPN, and the Markov model were employed in urine sediment classification. Wefound that the CNN algorithm (GoogLe-Net, ResNet, developed Alex-Net, LeNet-۵, VGG-۱۶,VGG-۱۹, yolov۳, R-CNN, Inception V۳, Xception) is applied most frequently (in ۳۳ studies).However, a deep learning model based on CNN’s Inception V۳ algorithm showed superior accuracy(۰.۹۹۴) comparatively for the detection of RBC, calcium oxalate, and bacteria in urine imagesamples.Conclusion: This paper gives a thorough analysis of the relative effectiveness of various traditionalmachine learning and deep learning algorithms for classifying urine sediments. Particularlydeep learning approaches require a larger number of training instances, which makes the processtime-consuming. Researchers can use this crucial comparative performance data to help themchoose the best machine learning algorithm for their project.

کلیدواژه ها:

Artificial intelligence ، Urine sediment ، Machine learning ، Urine analysis ، CNN ، Automatic classification