Tuesday, Aug 4: 10:30 AM - 12:20 PM
Contributed Posters
Thomas M. Menino Convention & Exhibition Center
Room: CC-Exhibit Hall A
Co Sponsors
Section on Statistics in Imaging
Presentations
Manual visual assessment of donor renal biopsy whole-slide images (WSIs) remains the current standard for predicting post-transplant renal function. Computer-extracted image features (pathomics) from these WSIs can enable more accurate prediction of kidney transplant recipient outcomes. While prior pathomics work has focused on single-site data, little attention has been given to building predictive models that leverage data from multiple institutions. Multi-site pathomics data exhibits batch effects and feature distribution shifts from differences in scanners and tissue processing, which can limit cross-site generalizability. We compared lasso, ridge, elastic net, random forest and XGBoost models for binary and continuous renal function outcomes using pooled cohorts with ComBat-based harmonization. Preliminary results indicate that ComBat-harmonized features improved test MSE for eGFR by up to 40% and DGF AUC by 23% on average across models compared to prior single-site studies. These findings suggest pathomics-based models can support clinical decision-making for kidney transplant allocation across institutions.
Keywords
Computational pathology
Data harmonization
Domain adaptation
Machine learning
Speaker
Ritesh Reddy Thipparthi, Department of Computer Science, University of Maryland, College Park, MD, USA
Co-Author(s)
Advay Monga, University of Maryland, College Park, MD, USA
Raymond Chen, Department of Computer Science, University of Maryland, College Park, MD, USA
Janelle Vo, School of Public Health, University of Maryland, College Park, MD, USA
Akshita Gupta, Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida
Anindya Paul, University of Florida
Luis Rodrigues, University Clinic of Nephrology, Faculty of Medicine, University of Coimbra, Coimbra, Portugal
Jeremy Rubin, Department of Epidemiology and Biostatistics, University of Maryland, College Park, MD, USA
Pinaki Sarder, University of Florida
The simple permutation-based inter-modal correspondence (SPICE) test is commonly used to test for intermodal correspondence, i.e., coupling between brain features obtained from different imaging modalities. It tests for the null hypothesis that the distribution of intermodal correspondence within a subject is no different than that between subjects. Implemented as a permutation test, SPICE assumes that subjects are exchangeable under the null. However, it is known that many neuroimaging findings vary with development, sex, or disease state. In this work, we set out to understand how additional covariates impact the results of SPICE. Using neuroimaging data from the Philadelphia Neurodevelopmental Cohort and simulated data, we identify a problem with SPICE analogous to "Simpson's paradox", wherein the results are confounded by group effects. We demonstrate how naive analyses may lead to confusing or improper results. We propose and evaluate two approaches for performing SPICE that account for measured confounding: (1) restricted permutations; (2) Fisher's method. Our results represent an important evaluation and extension of SPICE that increases its rigor and generalizability.
Keywords
Neuroimaging
Nonparametric
Hypothesis testing
Diabetic Foot Ulcers (DFUs) require frequent and accurate wound assessment to support clinical care and monitor healing. Traditional ruler-based measurements often involve direct contact with the wound surface, which can be invasive and increase infection risk.
We present DFU Toolkit, a web-based image analysis platform for end-to-end wound size measurement from real-world DFU images. The system accepts standard wound photographs, automatically segments ulcer boundaries, and estimates physical wound dimensions using a ruler scale embedded in the image, enabling non-contact measurement without manual tracing. Designed for practical use, DFU Toolkit provides an intuitive web interface that supports robust, reproducible wound measurement and facilitates scalable monitoring of DFU progression in both research and clinical settings.
Keywords
Diabetic foot ulcers
Wound size measurement
Medical image analysis
web-based toolkit
Pharmaceutical research and development often uses imaging of formulations as an essential component of their pipelines for small-molecule, biologic, and vaccine development. These imaging techniques include cryo-electron microscopy (cryoEM) for high-resolution imaging and micro-flow imaging (MFI) for analysis of sub-visible particles, among others. Deep learning approaches can provide automated, accurate and scalable analytics by performing a wide range of tasks, including object detection, segmentation, and classification. Benefits of quantitative imaging include enhanced quality control and improved scientific risk identification and mitigation. However, models often require modality and task-specific development and training. Foundation models for computer vision have emerged as promising tools to increase learning efficiency for a wide range of tasks by leveraging pre-training on large datasets. We show that using foundation models can result in improved performance in cryoEM segmentation and MFI particle classification, indicating that these models provide a flexible framework for generating insights from multiple imaging modalities to aid pharmaceutical development.
Keywords
biopharmaceutical
computer vision
imaging
Speaker
Hannah Horng, Merck & Co., Inc.
Co-Author(s)
Yueming Chen, Merck & Co., Inc.
Shubing Wang, Merck & Co., Inc.
Andy Liaw, Merck & Co., Inc.
Seema Irani, Merck & Co., Inc.
Caitlin Wood, Merck & Co., Inc.
Irene Chang, Merck & Co., Inc.
Daniel Skomski, Merck & Co., Inc.
Spatial context is essential for interpreting cell behavior in tissues, yet linking local microenvironments to cell phenotypes in Imaging Mass Cytometry (IMC) remains challenging. We present scalable methods to quantify two spatial feature sets around each cell. For cell–extracellular marker interactions, we aggregate pixel-level marker intensities within concentric rings surrounding each cell. For cell–cell interactions, we estimate the composition of neighboring cell types within increasing radii, enabling fixed-radius summaries and multi-scale interaction curves.
We use spatial indexing for fast queries of marker pixels and neighbors. Extracellular features are analyzed with mixed-effects models to estimate condition-specific marginal means while accounting for image-to-image variability. Cell–cell enrichment or depletion is assessed by comparing local neighborhood composition to overall image abundances across radii.
We implement these methods in the R package CytoHalo and demonstrate them on Collagen I–macrophage associations and macrophage-centered cell–cell interactions.
Keywords
IMC spatial analysis
Cell–cell interactions
Extracellular markers
Graph Neural Networks (GNNs) have shown promise in computer-aided diagnosis. However, the lack of interpretability hinders their clinical adoption. We present an explanation framework that adapts Local Interpretable Model-agnostic Explanations (LIME) to graph-level prediction by perturbing each patient's image set. Instead of superpixels, we sample images subsets (subgraphs) and estimate image-level influence on the model. To generate informative perturbations, we use a two-stage Adaptive Class-Balanced Sampling scheme. Stage I utilizes random sampling, and Stage II employs class-biased sampling, drawing 85% of images from the minority class pool to ensure an informative design matrix. Ridge and elastic net classifiers trained with 10-fold cross-validation quantify conditional image influence, and Pearson correlation provides complementary marginal relationships. We incorporate bootstrap-based uncertainty quantification to report standard errors and confidence intervals. To demonstrate the efficacy of our approach, we apply our method on 135 fatty liver disease patients from MacKay Memorial Hospital in Taiwan, offering clinicians intuitive visual explanations.
Keywords
explainable AI
graph neural networks
ultrasound imaging
interpretability
fatty liver
deep learning
Mediation analysis is often used in the behavioral sciences to investigate the role of intermediate variables that lie in the path between an exposure and an outcome. In recent years, there has been an increased interest in studying high-dimensional mediators (e.g., imaging and genomics data). In this talk, we describe an approach towards high-dimensional mediation that identifies latent low-dimensional representations of the set of potential mediators that maximize the indirect effect. We describe an estimation technique based on solving a generalized eigenvalue problem. We further explore links to canonical correlation analysis and partial least squares. The proposed methodology is flexible enough to work with multivariate scalar and functional data. We provide illustrations of both. The methods are applied to study how the relationship between maternal exposure and child behavioral outcomes is mediated by brain volume changes across the early years.
Keywords
Mediation analysis
Functional data analysis
Imaging
Color quantization is a technique used in Computer Imaging in which the number of true colors in an RGB image is reduced to a prespecified number of colors in a color palette, without appreciably reducing image quality. The importance of this technique lies in the fact that a color-quantized image can be displayed on devices or software platforms that are not fully capable of rendering all the colors of an image. Another important aspect of color quantization is that it facilitates image compression, owing to the limited colors in the color palette used to display images. In this work, we perform model-based color quantization that accurately captures the RGB channels of a two-dimensional digital image. Our model employs a Gaussian Copula Model with a Kumaraswamy Distribution as the marginal distribution to perform color quantization using the EM algorithm. In addition, we describe an updated model that incorporates spatial context into the color quantization process via penalization, providing an alternative approach. We test this model on a set of 2D digital images to assess the accuracy and performance for both the non-spatial and spatial component models.
Keywords
Gaussian copula model
Kumaraswamy distribution
Expectation-Maximization algorithm
Spatial component
K-means clustering
Visual-information-Fidelity (VIF)