Machine learning-assisted image label-free smartphone platform for rapid segmentation and robust multi-urinalysis.

Abstract

This study presents a groundbreaking approach for the early detection of chronic kidney disease (CKD) and other urological disorders through an image-label-free, multi-dipstick identification method, eliminating the need for complex machinery, label libraries, or preset coordinates. Our research successfully identified reaction pads on 187 multi-dipsticks, each with 11 pads, leveraging machine learning algorithms trained on human urine data. This technique aims to surpass traditional colourimetric methods and concentration-colour curve fitting, offering more robust and precise community screening and home monitoring capabilities. The developed algorithms enhance the generalizability of machine learning models by extracting primary colours and correcting urine colours on each reaction pad. This method's cost-effectiveness and portability are significant, as it requires no additional equipment beyond a standard smartphone. The system's performance rivals professional medical equipment without auxiliary lighting or flash under regular indoor light conditions, effectively managing false positives and negatives across various categories with remarkable accuracy. In a controlled experimental setting, we found that random forest algorithms, based on a Bagging strategy and applied in the HSV colour space, showed optimal results in smartphone-assisted urinalysis. This study also introduces a novel urine colour correction method, significantly improving machine learning model performance. Additionally, ISO parameters were identified as crucial factors influencing the accuracy of smartphone-based urinalysis in the absence of additional lighting or optical configurations, highlighting the potential of this technology in low-resource settings.