Deep-learning-based color image segmentation for biological application

Abstract

Detection of proteinuria in urine is an early sign of the progression of cardiovascular and chronic kidney disease. In general, colorimetry detection in urine dipsticks is carried out for a screening of proteinuria. Although low sensitivity and vision assessment has hindered its utility. Smartphone-based colorimetry detection using a urine dipstick plays an important role to estimate the concentration of albumin. Segmentation of albumin patches in urine dipstick is the first step for the automatic estimation of albumin concentration in the urine sample. Here, we present a deep learning-based object detection technique i.e. You Only Look Once (YOLOv5), for the segmentation of albumin patches from urine dipstick (Uristix Siemens) for the estimation of albumin concentration. A comparison between different segmentation techniques to accurately arrive at the region of interest has been demonstrated. Four different smartphones i.e. iPhone SE, Realme C11, Redmi 8A Dual, and Samsung Galaxy M01 were used to capture the images of the urine dipstick, at six different illumination conditions i. e. 500 Lux, 400 Lux, 300 Lux, 200 Lux, 100 Lux, and 50 Lux. The training was done using iPhone SE with eight different albumin concentrations: 10 mg/L, 20 mg/L, 40 mg/L, 80 mg/L, 160 mg/L, 320 mg/L, 640 mg/L, and 1280 mg/L. The proposed model is fast and robust and it helps minimize the effect of different light conditions and color variations across smartphone models.