Graduation Semester and Year
Spring 2026
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Engineering
Department
Computer Science and Engineering
First Advisor
Jacob M. Luber
Abstract
Hematoxylin and eosin (H&E) staining remains central to cancer diagnosis, providing morphological information essential for pathological assessment. Immunohistochemistry (IHC) and newer multiplexed imaging technologies complement H&E by revealing molecular information critical for accurate tumor subtyping and treatment decisions. In practice, however, H&E and IHC are obtained from different consecutive sections that are not spatially aligned, comprehensive multiplexed panels are expensive and tissue-consumptive, and not all stains are available at every clinical site, limiting comprehensive molecular profiling and the full diagnostic potential of these technologies in clinical practice. This dissertation addresses these gaps through three complementary generative deep learning studies in computational pathology.
The first study introduces an SSIM-guided conditional GAN for imputing held-out protein expression channels from a reduced acquired subset in multiplexed spatial proteomics data. Using CODEX/PhenoCycler data with up to 29 markers, SSIM-based hierarchical clustering identifies informative channel sets, and scalability analysis demonstrates potential for learning inter-biomarker relationships across larger panels.
The second study investigates cross-domain synthesis from multiplex immunofluorescence (mIF) to H&E, bridging molecular and morphological imaging. A proposed multi-level Vector-Quantized Generative Adversarial Network (VQGAN) outperformed conditional GAN baselines on two datasets of varying channel configurations (18 and 58 channels), demonstrating superior performance in both image reconstruction and downstream tasks including nuclei segmentation and tissue classification.
The third study presents UNIStainNet, a foundation-model-guided architecture for virtual immunohistochemistry staining from routine H&E sections. By conditioning a SPADE-UNet generator on dense spatial tokens from a frozen pathology foundation model (UNI) and employing a misalignment-aware loss suite, a single unified model achieves state-of-the-art performance across multiple IHC markers (HER2, Ki67, ER, PR) while matching per-stain specialists at one-quarter the parameter count.
Together, these studies establish a bidirectional virtual staining pipeline spanning within-domain multiplexing, molecular-to-morphological translation, and morphological-to-molecular generation, advancing generative artificial intelligence in computational pathology.
Keywords
Computational pathology, generative artificial intelligence, virtual staining, image-to-image translation, biomarker prediction, medical image synthesis, digital pathology
Disciplines
Computational Engineering | Computer Engineering
License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Saurav, Md Jillur Rahman, "Generative Imaging for Computational Pathology" (2026). Computer Science and Engineering Dissertations. 13.
https://mavmatrix.uta.edu/cse_dissertations2/13