Chronic kidney disease is the ninth leading cause of death in the United States. The current process for pathological diagnosis involves pathologists manually reviewing histochemically stained tissue slides. This analysis is also used to inform further treatment and is therefore critical to patient outcomes. In this project, we aim to improve the robustness of the diagnostic process by utilizing machine learning models to identify and classify features indicative of kidney disease on whole slide images. We manually annotate Masson’s Trichrome stained kidney tissue images from our collaborators at the University of Illinois for three functional structures (tubular cytoplasm, tubular basement membrane, glomerulus) and three indicators of damage to the kidney (fibrosis, edema, and inflammation). These annotations are used to train our VGG16 convolutional neural network model to classify patches of unmarked whole slide images into the four categories: tubular cytoplasm, fibrosis, inflammation, and glomerulus. We also address data variability that often comes from differences in the histochemical staining procedure across labs resulting in inconsistency across stains/imaging that can typically affect the generalizability of deep learning models. To address this, we are training a CycleGAN for image-to-image translation as a method of stain normalization and investigating the effect on the accuracy of our VGG16 model. Additionally, I will be training a model to identify the cortex versus medulla regions of the kidney to add to the pipeline for area-specific evaluations. Our research with integrating machine learning models within renal pathology aims to decrease the time and manual labor needed in the process and increase the accuracy of diagnoses.

 Immune cells make up the body's defense against cancer and observing their spatial distribution in a tumor can provide information about patient prognosis. However, it is difficult and time consuming to identify each immune cell in images from cancer biopsies in order to perform spatial analysis. Additionally, stained immune cells are hard to distinguish by appearance in unprocessed multispectral images due to the overlapping or "mixing" of signals coming from different channels. A computational tool could efficiently identify the immune cells in a tumor. The goal of this project is to build a digital pipeline to identify each immune cell in a multispectral image of a tumor and make it generalizable to multispectral images from any source. First, we use an unsupervised method to break up mixed multispectral images into clusters. The user selects a subset of clusters that do a good job of isolating each type of immune cell. A sample of these user-selected results are used to train a supervised machine learning model. The trained model assigns a label to each cluster to classify the entire image. Preliminary results have shown that clusters can usually be assigned to the correct label with over 50% certainty. We anticipate that the clusters will show good agreement with clinician classifications. This pipeline will allow for immune cell identification with less human involvement than pathologist annotation and without requiring spectral unmixing, a preprocessing step that typically takes hours. In the future, we will test this pipeline with varying amounts of training data coming from different sources and integrate it with spatial analysis to capture immune signatures of disease.