My research focuses on developing computational and machine learning methods for analyzing complex, multimodal biomedical data, with applications in cancer genomics, medical imaging, and AI-driven diagnostics.
Research Projects
Graph-Based and Multimodal Learning for Single-Cell and Spatial Omics Analysis
This research develops advanced computational methods for analyzing single-cell and spatial transcriptomics data, with a focus on scalable and biologically informed clustering and representation learning. By integrating graph neural networks, self-supervised and contrastive learning, and prior biological interaction networks, these approaches capture complex cellular relationships and improve cell-type identification. The work further extends to multiomics integration and multimodal modeling with histology images, enabling more accurate, spatially aware characterization of cellular environments in complex tissues.
Multimodal Graph and Transformer Learning for Integrative Cancer Analysis
This research develops a unified computational framework for integrating genomics and histopathology data using advanced machine learning techniques. It combines attention-based graph neural networks, transformer-based multiple instance learning, and contrastive learning to extract meaningful representations from multi-omics data and whole-slide images. By incorporating prior biological knowledge and modeling spatial and structural relationships, these methods enable more accurate and interpretable cancer subtype classification, survival prediction, and phenotype modeling from large-scale biomedical datasets.
Morphology-Aware Graph Learning for Weakly Supervised Analysis of Histopathology Images
This work develops advanced weakly supervised deep learning frameworks for analyzing whole-slide histopathology images, addressing challenges such as tissue heterogeneity and limited annotations. Our research focuses on iterative patch refinement and sampling, graph-based modeling of spatial tissue structure, and cross-modal semantic alignment using clinical concepts and language models to capture diagnostically relevant patterns. Our goal is to produce more accurate, interpretable, and spatially coherent phenotype predictions, enabling reliable and clinically meaningful cancer grading, by combining transformers, graph neural networks, and morphology-aware learning.
Longitudinal and Multimodal AI-CAD for Mammogram Analysis
This research develops advanced AI methods for breast cancer detection using longitudinal mammograms and 3D tomosynthesis data,mimicking how radiologists compare prior and current scans. The proposed approaches integrate Siamese networks, transformer-based models, graph neural networks, and correlation-aware learning to capture subtle temporal changes and complex tissue structures. By leveraging multi-scale features, novel distance functions, and weakly supervised learning, these methods improve detection, classification, and weakly supervised segmentation of breast abnormalities to enable more accurate and reliable computer-aided diagnosis.
Generative Modeling of Longitudinal Tumor Development in Mammography
This research introduces a novel generative AI framework for synthesizing realistic breast tumors in longitudinal mammograms, addressing the scarcity of annotated temporal data. By leveraging Transformer-based encoders, variational latent modeling, and attention-driven decoding, the approach captures both anatomical structure and temporal progression between prior and current scans. The framework produces anatomically consistent and diverse tumor patterns, enabling improved data augmentation, temporal modeling, and development of more robust AI systems for breast cancer detection.
Few-Shot Domain-Incremental Continual Learning for Adaptive Medical Imaging Models
This work develops a few-shot domain-incremental continual learning framework for updating medical imaging models under evolving data distributions. The approach enables models to adapt to new clinical conditions using limited new data while preserving prior knowledge to avoid catastrophic forgetting. By integrating selective parameter updating, memory-efficient replay, correlation-aware regularization, and expert-in-the-loop feedback, the framework supports stable and incremental model refinement. A teacher–student architecture further ensures robust, well-calibrated predictions while leveraging temporal information and domain shifts, enabling practical and sustainable deployment of AI systems in real-world clinical environments.
Quantum-Inspired representation Learning for Robust Longitudinal Mammogram Analysis
This research develops a novel quantum-information-inspired framework for modeling similarity in longitudinal mammogram analysis, addressing limitations of traditional distance functions that fail to capture complex, correlated tissue changes. By interpreting deep features as statistical states and introducing fidelity-based and correlation-aware distance measures, the approach captures nonlinear relationships between prior and current images. Combined with geometry-aware regularization and robust training strategies, these methods improve representation quality, robustness to distribution shifts, and diagnostic accuracy, enabling more reliable and interpretable AI systems for medical imaging.
Uncertainty-Aware 3D Reconstruction and Registration in Multiplex Microscopy Imaging
This research develops advanced computational frameworks for 3D reconstruction and registration of biological tissue from serial 2D multiplex and microscopy images, addressing noise, distortion, and alignment challenges. The proposed methods integrate Bayesian transformers, uncertainty-aware feature learning, graph-based cell tracking, and hybrid deep learning–based registration techniques to accurately align and link cells across sections. By combining robust feature extraction, adaptive outlier handling, and spatial consistency modeling, these approaches enable reliable reconstruction of volumetric tissue architecture and improve precision in complex biomedical imaging analysis.