A deep-learning application for epithelial cells image detection

Abstract

Urinary particles are used to evaluate the different urinary tract diseases in patients. Currently, doctors use the traditional methods for urinalysis such as urine dipstick, urine culture and microscopy. Microscopy is an effective method for the diagnosis and treatment of many kidney and urinary tract diseases. However, manual microscopic examination of urine is labor-intensive, subjective, imprecise, and time-consuming. In this project, we proposed the development of a different deep learning models classifier for an automated microscopic urinalysis system for epithelial cells. A dataset was constructed from scratch taking urine samples from the Hospital Ginequito obtaining a total of 857 images. Then, the images were labeled into urine samples with and without epithelial cells for binary classification. Last, we created three deep learning models using the InceptionV3 architectures with different series of fully connected layers randomly initialized and ReLU activation, a dropout rate of 0.2 and a final sigmoid layer for classification. The best model obtained a training accuracy of 81.89% with sensitivity of 77.84%, specificity of 85.94% and precision of 84.70% and a validation accuracy of 84.28% with a sensitivity of 87.50%, specificity of 81.25% and precision of 82.35%. It was concluded that microscopic urinalysis can be done automatically, this opens the door for the classification of more urine particles with improved metrics.