Mitigation of Colorimetric Variabilities in Smartphone-Assisted Diagnostic Reader Using Domain-Adaptation

Urinary albumin is an excellent marker for the diagnosis of chronic kidney disease. Smartphone assisted accessory-free analyzers are gaining popularity in the quantification of urinary albumin in point-of-care settings, though, presently, they are widely used as a screening tool. Application of smartphone-based colorimetry systems are limited because they suffer from several operating problems such as camera settings and illumination conditions, that can change the colorimetric values. In addition, they suffer from the problem of domain shift i.e. extreme performance degradation when tested data (source) was captured in a setting different from that of the training (target) data. In this work, the problem introduced due to smartphone camera settings and illumination conditions, has been addressed by applying a domain adaptation deep learning method. It is the amalgamation of a generative model and convolutional neural network. Images captured using an iPhone under 3500K light conditions were used as the target dataset and other domains including different smartphones and light conditions were utilized as source datasets. Given test data from the source, its ‘closest clone’ was derived using generative model-based Pix2pixGAN and mapped to the target data. Further, a customized CNN model was used for the classification of the closest clone data. The proposed method yields an accuracy of ~88% for inter-phone repeatability on test data. The efficacy of the proposed model was also evaluated under different illumination conditions.