Unsupervised Machine Learning to Identify Important Features for Focal Segmental Glomerulosclerosis
Abstract
Focal segmental glomerulosclerosis (FSGS) is one of the leading causes of kidney disease worldwide. FSGS causes serious scarring that leads to permanent kidney damage and is one of the causes of a serious condition known as Nephrotic Syndrome. It is the most common primary glomerular disorder causing end-stage renal disease in the United States. However, pathologists use visual classification of glomerular lesions to assess patient samples when diagnosing kidney disease, and the assessment is qualitative, or semi-quantitative at best, and reproducibility is challenging. To overcome substantial interobserver variability, computer-aided algorithms such as machine learning/deep learning can help provide an objective and reliable method of assessment. Recently, digital pathology has begun to play an important role in modern clinical practice with the growing availability of whole slide digital scanners, increasing computing power, faster networks, and bigger storage spaces. In addition, several studies have investigated the application of artificial intelligence (AI) in digital pathology. However, AI has been mostly used for image-based diagnosis in radiology and oncology. Image segmentation and classification using various computer-aided methods, especially machine learning/deep learning, have considerably expanded the scope of medical image analysis. This supervised analysis, measuring known structures or features believed to be relevant, has also been successfully employed for a number of biomedical analyses. Unbiased (unsupervised) image analysis, however, provides several advantages for research as it has the potential to capture biologically meaningful information from the complexity of an image and provide greater reproducibility and reliability. Our proposal seeks to address FY21 PRMRP focus areas concerned with improving our understanding of FSGS disease heterogeneity by applying digital pathology and computer-assisted image analysis approaches to kidney biopsies in addition to molecular profiling with transcriptomics. In this study, we will develop a methodology for image feature extraction based on the bag-of-words (BoWs) approach and apply to NEPTUNE FSGS cases to find previously unknown features that predict patient outcomes. The BoW model is originally from information retrieval and natural language processing and is commonly used in document classification tasks where the frequency or occurrence of each word is used as a feature for training a classifier. We hypothesize that unsupervised machine learning algorithms without human’s input can find previously unknown features that predict patient outcomes and are associated with gene expression profiling. To do this we propose to: Aim 1: To determine if quantitative features from an unsupervised machine learning with BoWs approach predicts patient outcomes. Aim 2: To determine transcriptomic determinants that are associated with histopathology features identified by unsupervised machine learning. By using computer-aided and quantitative methods to identify important histopathology features regarding the patient outcomes in FSGS, and to correlate identified features with gene expression profiles, we expect that this work will result in new clinically predictive categorization of FSGS. The identification of important histopathology features will inform future studies that aim to detect these histopathology features through computer automation. Further, we will understand how FSGS heterogeneity impacts risk of disease presentation, clinical course, patient outcomes, and FSGS mechanisms.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 28, 2022
- Source ID
- W81XWH2210032
Entities
People
- Joonsang Lee
Organizations
- United States Army
- University of Michigan