Computational Classification Methods in Renal Pathology

Abstract

<p>When assessing kidney biopsies, pathologists use Light microscopy (LM), immunofluorescence, and electron microscopy to describe and diagnose glomerular lesions and diseases. These are laborious, costly, fraught with inter-observer variability, and can have delays in turn-around time. Thus, computational approaches can be designed as screening and/or diagnostic tools, potentially relieving pathologist time, healthcare resources, while also having the ability to identify novel biomarkers, including subvisual features.</p> <p>This thesis compares a novel Biomarker feature extraction (BFE) model and test several pre- trained deep learning models to diagnose three glomerular diseases using Periodic acid-Schiff (PAS) stained images. The BFE model extracted a panel of 233 explainable biomarkers related to colour, morphology, and microstructural texture that describe underlying pathology. These biomarkers were subsequently narrowed to ten morphological and microstructural texture fea- tures for classification. The proposed pre-trained deep learning models used data augmentation and Gradient-weighted Class Activation Mapping (Grad-CAM) for better performance and inter- pretability. All classification models were then used to classify Minimal change disease (MCD), Membranous nephropathy (MN), and Thin-basement membrane nephropathy (TBMN) diseases on a glomerular and patient-level basis.</p> <p>The BFE model resulted in a glomerular validation accuracy of 67.6% and testing accuracy of 76.8%. All deep learning approaches had higher validation accuracies (most for VGG16 at 78.5%) but lower testing accuracies. The highest testing accuracy at the glomerular level was VGG16 at 71.9%, while that at the patient-level was InceptionV3 at 73.3%. Patient-level classification results showed higher accuracy for the BFE model (86.76%) compared to all deep learning methods.</p> <p>The BFE model was superior to three established deep learning models. The small dataset was likely contributory to the lower performance of the deep learning models, while model in- terpretability was clearly superior for the BFE. The results highlight the technical performance of both the traditional and deep learning approaches and describe how they may translate to clinical utility. Furthermore, these algorithms can be applied to clinical datasets for novel prognostic and mechanistic biomarker discovery.</p>