PROJECT SUMMARY
Our overarching goal is to enable the AI-empowered 3D histopathological interpretation on routine digitized renal
tissue biopsies, so as to (1) allow renal pathologists to perform a reproducible 3D phenotyping on serial 2D whole
slide images (WSI), (2) advance the characterization of kidney-allograft rejection phenotypes on kidney
transplant patients with cutting-edge 3D computer vision, and (3) equip clinical scientists with an advanced 3D
spatial transcriptomics analytics tool to investigate the anatomical-molecular associated causes of chronic kidney
disease (CKD). Our novel 3D histopathological interpretation, with 3D computer vision (Map3D toolkit) and
3D spatial transcriptomics, will open a new door for performing reproducible clinical phenotyping (Pheno3D
toolkit), identifying and validating new 3D imaging and molecular biomarkers (GPS3D toolkit), and ultimately
advancing the patient care with personalized diagnosis and prognosis options for a wide range of CKD.
Despite more than 25 years of exploitation of digital pathology, the presence, significance and
characteristics of 3D contextual information in renal histopathological assessment have been largely
overlooked. The current 2D interpretation on renal histopathology is error-prone and less reproducible due to
the heterogeneity of tissue morphologies (e.g., glomeruli, tubules, vessels) across 3D serial sections. For
example, our previous study on segmental glomerulosclerosis (GS) in patients with nephrotic syndrome and
idiopathic FSGS, the percent of GS increased from 31.5 +/- 6.8% to 48.0 +/- 6.6% (P < 0.025) in adults by
replacing a 2D single section analysis with 3D serial section analysis. Moreover, 2D based phenotyping can also
hinder the discovery of new biomarkers via state-of-the-art spatial transcriptomic techniques. As an example, a
glomerulus with focal segmental glomerulosclerosis (FSGS) can have a normal appearance on a specific 2D
section, which might lead to an opposite molecular finding using 2D spatial transcriptomics
The core tenant of this proposal is NOT developing a new 3D imaging modality, but rather, to develop
technologies that enable reproducible 3D characterization on routine 2D renal histopathological biopsies (with
trivial added cost), so as to advance the care of future patients with renal diseases. To this end, we will:
Aim 1. Develop novel 3D computer vision tools (Map3D) to facilitate renal pathologists in modeling,
quantifying, and visualizing 3D renal histopathological tissues from routine 2D digital histopathology.
Impact: Allow renal pathologists to perform a reproducible 3D phenotyping on serial 2D whole slide images (WSI).
Aim 2. Develop 3D phenotyping tools (Pheno3D) to advance the characterization of kidney-allograft
rejection for kidney transplant patients via 3D computer vision and self-supervised deep learning.
Impact: Advance the characterization of kidney-allograft rejection phenotypes for kidney transplant patients.
Aim 3. Develop 3D computer vision algorithms for 2D and 3D spatial transcriptomics (GPS3D toolkit).
Impact: Equip clinical scientists an 3D spatial transcriptomics analytics tool to investigate image-omics interaction.
