Section on Statistics in Imaging Contributed Session 1

Dan Rowe Chair
Marquette University
 
Sunday, Aug 3: 2:00 PM - 3:50 PM
4013 
Contributed Papers 
Music City Center 
Room: CC-101B 

Main Sponsor

Section on Statistics in Imaging

Presentations

A Novel Infimum Dimension Nash Embedding for 3D Projective Shape Data

An isometric embedding of a Riemannian manifold in Euclidean space is an embedding such that the Riemannian structure on each tangent space is induced by the scalar product in the ambient space; such an embedding preserves path lengths, angles, areas and volumes. Such a mapping preserves much of the geometric information contained in the original Riemannian manifold where the data is valued allowing for some computationally advantages. In addition if the Riemmanian manifold is homogeneous then we would also want the such an embedding, to be equivariant. The Veronese-Whitney (VW) embedding is an isometric equivariant embedding for 3D projective shape data. It is in fact the only embedding used in the analysis of 3D projective shape data, at the present time. In this paper we consider a novel equivariant isometric embedding, the Nash embedding into a Euclidean space of a lower dimension than the Euclidean space where the Veronese-Whitney embedding is valued . We compare the performance of the novel Nash embedding-based statistical techniques with those  based on the VW embedding in Monte Carlo studies and a real data example. 

Keywords

Extrinsic shape analysis

Nash embedding

Nonparametric test

Three-dimensional projective shape analysis 

Co-Author(s)

Vic Patrangenaru, Florida State Univerity
Mihaela Pricop-Jeckstadt, Universitatea Națională de Știință și Tehnologie Politehnica București

First Author

Robert Paige, Missouri S&T

Presenting Author

Robert Paige, Missouri S&T

Absolute risk of cognitive impairment by Alzheimer's Disease amyloid stages

Despite advances in Alzheimer's disease (AD) research, limited information exists regarding the absolute risk of mild cognitive impairment (MCI) in cognitively unimpaired (CU) individuals with abnormal AD biomarkers, particularly when accounting for competing risks of death. We included 5,858 participants from the Mayo Clinic Study of Aging (MCSA) to evaluate AD amyloid stage as a predictor of clinical progression to MCI or dementia. The data includes long-term follow-up information on death and dementia beyond active study participation, which mitigates potential bias due to dropout. We predicted 10-year and lifetime risks of MCI and dementia, accounting for the competing risks of death, given amyloid PET stages, sex, APOE4 status, and baseline age. Results are based on a hidden Markov model. AD amyloid staging based on amyloid PET Centiloid values was the strongest predictor of lifetime risk for MCI or for dementia. Higher Centiloid levels amplified age effects on the risk of MCI, whereas for dementia, amyloid stage effects surpassed age effects. For 10-year risk, age was the dominant factor, whereas for lifetime risk, amyloid stages had a greater influence. 

Keywords

Hidden Markov Models

Alzheimer's Disease

Absolute Risk 

Co-Author(s)

Terry Therneau, Mayo Clinic
Clifford Jack, Mayo Clinic
Heather Wiste, Mayo Clinic
Prashanthi Vemuri, Mayo Clinic
Jon Graff-Radford, Mayo Clinic
Ronald Petersen, Mayo Clinic
Dave Knopman, Mayo Clinic
Val Lowe, Mayo Clinic

First Author

Mingzhao Hu, Mayo Clinic

Presenting Author

Mingzhao Hu, Mayo Clinic

Assessing Topological Associations Between Immune Cell Structures in Spatial Proteomic Imaging Data

Spatial proteomic technologies reveal immune cells organization, offering critical information about immune function and disease mechanisms. Standard methods for assessing immune cell interactions rely on simplistic summary statistics that fail to accommodate variations in scan orientation, cell count, and location variations across individuals, thereby do not directly evaluate spatial structures.

To address this, we use topological data analysis (TDA) with persistent homology (PH) to capture spatial structure. PH systematically translates spatial information of cells into topological summaries, producing k*n persistence diagrams-one for each cell type per individual. Pairwise L1 distances between persistence diagrams for individual cell types form k n*n distance matrices that capture structural differences across the population. The Kernel RV is then used to identify associations between the spatial structures of different immune cell types. Simulations and real data analyses show our approach is often more powerful, particularly at assessing global structures, while still protecting type I error, serving as a powerful new approach investigating spatial immune cell interaction. 

Keywords

Spatial Proteomic Imaging Data

Persistent Homology

Kernel RV

Topological Data Analysis 

Co-Author(s)

Michael Wu, Fred Hutchinson Cancer Center
Sarah Samorodnitsky, Fred Hutch Cancer Research Center

First Author

Jingyi Guan

Presenting Author

Jingyi Guan

Estimating Temporal Evolution of Topological Features in Image Data

One popular technique in Topological Data Analysis (TDA) called persistent homology (PH) is used to describe holes in an image through their dimension and the functional values (e.g., thresholds or scales) at which they are created and filled. While TDA has been successfully applied to identifying shape in a static image through its hole structure, estimating the changes in that hole structure within a time-evolving image set is relatively understudied. We develop a method which first identifies statistically significant topological features in the spatial and temporal dimensions simultaneously. These higher-dimensional topological features are then used to establish temporal connections between the lower-dimensional features they are built from, effectively separating spatial connections from temporal connections. The spatial structure of the lower-dimensional features can be analyzed at each time-point separately and their temporal evolution represented on a ZigZag diagram (topological summary statistic focused on time dynamics). The method's effectiveness in capturing the emergence and progression of topological features is tested on a time series of images from cell wounds. 

Keywords

Image Processing

, Spatiotemporal Analysis

Topological Data Analysis

High Dimensional Statistics 

Co-Author(s)

Jessi Cisewski-Kehe, University of Wisconsin-Madison
Jun Zhu, University of Wisconsin - Madison
William Bement, University of Wisconsin

First Author

Susan Glenn, Los Alamos National Laboratory

Presenting Author

Susan Glenn, Los Alamos National Laboratory

Optimal Generalized Gaussian Scale Estimator with Applications to Lloyd-Max Quantization of Images

The generalized Gaussian distribution (GGD) is a versatile parametric family extensively used in signal and image processing for its ability to model diverse data distributions. In image compression algorithms like JPEG, the distributions of discrete cosine transform coefficients for a broad range of images are often represented by the GGD, encompassing widely used distributions such as Laplace and Gaussian. While extensive research focuses on estimating the GGD's shape parameter, fewer studies have developed accurate and optimal approaches for estimating the scale parameter, which is critical for controlling distribution spread and compression performance. We propose a novel optimal estimator for the GGD scale parameter and derive its exact mean squared error (MSE). We show analytically that our estimator has a uniformly smaller MSE than that of the maximum likelihood estimator. When applied to Lloyd-Max quantization of real images, our estimator demonstrates excellent performance, balancing feature preservation, compression efficiency, and minimal distortion. 

Keywords

Optimal Scale Estimator

Generalized Gaussian Distribution

Maximum Likelihood Estimator

Mean Squared Error

Image Compression

Lloyd-Max Quantization 

First Author

Kai-Sheng Song, University of North Texas

Presenting Author

Kai-Sheng Song, University of North Texas

Outcome Prediction using Image Features with Conformal Quantile Regression

Estimated glomerular filtration rate (eGFR) is a continuous biomarker of kidney function and an important clinical outcome in glomerular and other kidney diseases. The use of demographic, clinical, and kidney biopsy image data to predict future eGFR and quantify prediction uncertainty is crucial for risk stratification and clinical decision-making. Conformal quantile regression (CQR) provides a statistical framework to estimate prediction intervals around continuous outcomes with statistical guarantees about coverage of the true outcomes. However, CQR has not been explored in the context of predicting eGFR using image data that include generated regressors. In this study, we conducted a simulation study to test the performance of CQR in constructing prediction intervals of continuous outcomes from generated regressors. We demonstrated that CQR is robust to additive measurement error in the generated regressors but large samples are required for optimal coverage with functionally misspecified regressors. Finally, we used real-world glomerular disease kidney biopsy image features to predict eGFRs and demonstrated that CQR prediction intervals provide reliable coverage in real data. 

Keywords

Computational pathology

Conformal quantile regression

Image features

Generated regressors

Prediction intervals 

Co-Author(s)

Larry Han, Northeastern University
Jarcy Zee, University of Pennsylvania

First Author

Jeremy Rubin, University of Pennsylvania

Presenting Author

Jeremy Rubin, University of Pennsylvania