Urine analysis is used by clinicians to diagnose infections, kidney function, conditions like pregnancy and diabetes, by examining microscopic samples to detect and classify clinically significant objects. These objects are most commonly red blood cells (RBCs) or white blood cells (WBCs), epithelial-casts, crystals, microorganisms like bacteria, yeast, and artifacts. Artifacts in urine samples occur due to multiple reasons, ranging from fibers and dust being present in the sample to blurring of an out-of-focus object. The growing availability of high-resolution images of biological samples has made pathology and microscopy a popular application area for deep learning techniques.

In building robust convolutional neural network based automated urine sample analyzers, multiple artifacts present in urine microscopy images need to be identified and rejected by these systems during test time, while correctly classifying clinically significant objects. We tackle this problem by building state-of-art models trained on a microscopic urine sample image data-set provided by Sigtuple, that adapt existing deep neural network architectures to support identification and rejection of artifacts belonging to unseen/unknown objects.