Computing PhD Research

Scientific Computing and Imaging Institute

Kalhert School of Computing

University of Utah

Overview

My dissertation was in machine learning for medical image analysis, with a focus on Statistical Shape Modeling (SSM). SSM enables population-based quantitative analysis of anatomical shapes such as bones and organs, informing clinical diagnosis. However, traditional optimization-based SSM generation techniques have limitations that hinder their widespread adoption. My work made SSM more feasible, trustworthy, and broadly applicable via three research aims.

Aim 1: Uncertainty Quantification in SSM from Images

Deep learning frameworks, such as DeepSSM, extract SSM directly from unsegmented images (such as CT or MRI scans) with little manual overhead. While such approaches alleviate the time-consuming and expert-driven workflow of traditional SSM generation techniques in inference, they can produce overconfident estimates of shape that cannot be blindly assumed to be accurate. In my work, I have adapted such networks to convey what they do not know via granular estimates of uncertainty. This is critical in sensitive clinical applications.

Uncertain-DeepSSM: From Images to Probabilistic Shape Models

ShapeMI Workshop at MICCAI 2020 (Best Paper Runner-Up)

Uncertain-DeepSSM_Graphical_Abrstract-v3.pdf

From Images to Probabilistic Anatomical Shapes: A Deep Variational Bottleneck Approach

MICCAI 2022 (Early Accept)

VIB_poster.pdf

Fully Bayesian VIB-DeepSSM

MICCAI 2023 (Early Accept)

BVIB_poster.pdf

Benchmarking Scalable Epistemic Uncertainty Quantification in Organ Segmentation

UNSURE Workshop at MICCAI 2023

Unsure_poster.pdf

Weakly Supervised Bayesian Shape Modeling from Unsegmented Medical Images

ShapeMI Workshop at MICCAI 2024 (Oral Presentation)

Aim 2: Learning Spatiotemporal SSM

Numerous clinical investigations require understanding changes in anatomical shape over time, such as in dynamic organ cycle characterization or longitudinal analyses (e.g., for disease progression). Spatiotemporal statistical shape modeling (SSM) allows for quantifying and evaluating dynamic shape variation with respect to a cohort or population of interest. Existing SSM generation approaches assume sample independence and thus are unsuitable for sequential dynamic shape observations. I've extended the optimization scheme to capture a statistically significant time dependency directly.

Spatiotemporal Cardiac Statistical Shape Modeling: A Data-Driven Approach

STACOM Workshop at MICCAI 2022

4D_STACOM22_Poster.pdf

Learning Spatiotemporal Statistical Shape Models for Non-Linear Dynamic Anatomies

Frontiers in Bioengineering and Biotechnology Journal 2023

Aim 3: SSM from Point Clouds

Traditional methods for creating SSMs need noise-free surface meshes or binary volumes, rely on assumptions or predefined templates, and simultaneously optimize across the entire cohort, leading to lengthy inference times given new data. My work overcomes these barriers by providing a data-driven solution that infers SSMs directly from raw point clouds, reducing inference burdens and increasing applicability as point clouds are more easily acquired.

Google Sites

Report abuse