Research

Overview

My research program advances the frontiers of Artificial Intelligence by developing novel, mathematically grounded deep learning methodologies for complex, multi-modal, and heterogeneous data . I address fundamental challenges in AI—specifically in disentangled representation learning, graph neural networks (GNNs), and vision-language alignment—to create systems that are not only accurate but also interpretable and robust to the noise inherent in biological systems.

While my innovations are rooted in core AI algorithms, I apply them to solve high-stakes problems in computational neuroscience and medical imaging, moving from "methods-to-impact.".

The DiA architecture: A Vision-Language Mixture-of-Experts VAE that disentangles modality-specific and shared latents to drive a LLaMA-X decoder.

1. Robustness through Disentangled Generative AI

DiA-gnostic VLVAE: Disentangled Alignment for Robust Radiology Reporting (AAAI 2026 Oral Presentation)

A critical challenge in multi-modal medical AI is "feature entanglement," where shared disease markers are conflated with modality-specific noise, leading to model failure when clinical context is missing . We developed DiA-gnostic VLVAE, a novel probabilistic framework that learns a tri-factor latent space explicitly separating vision-specific, language-specific, and shared semantics .

Methodology: We utilize a Mixture-of-Experts (MoE) posterior to infer shared latents, ensuring robustness even when modalities are incomplete.
Innovation: A custom Disentangled Alignment constraint enforces orthogonality between specific factors while maximizing contrastive consistency in the shared space.
Impact: The model achieves state-of-the-art performance in radiology report generation on IU X-Ray and MIMIC-CXR, even in "missing modality" scenarios.

2. Physics-Guided & Explainable Graph Learning

NeuroKoop: Neural Koopman Operators for Brain Dynamics (BHI 2025 NSF-EMBS-Google Young Professional NextGen Scholar Award)

Standard deep learning often treats brain connectivity as static. To capture the underlying dynamics of neurodevelopment, we introduced NeuroKoop, a graph neural network that integrates structural (SBM) and functional (FNC) connectomes using Koopman operator theory.

Methodology: Instead of static fusion, we lift graph snapshots into a linear dynamical latent space using a learnable Neural Koopman operator.
Innovation: This allows us to model the "trajectory" of brain network evolution, conditioned on cognitive scores (e.g., Working Memory).
Impact: The model significantly outperforms traditional baselines in identifying prenatal drug exposure effects in the massive ABCD dataset.

NeuroKoop Overview: Dynamic latent space fusion of structural and functional connectomes driven by a Neural Koopman operator.

Workflow for augmenting Deep Ultraviolet (DUV) datasets using Diffusion Probabilistic Models to improve intraoperative breast cancer margin assessment.

3. Translational AI: Data Synthesis & Rare Disease

Diffusion Probabilistic Models for Intraoperative Cancer Assessment (NIH R01EB033806 Funded Research)

Data scarcity is a major bottleneck in developing AI for novel imaging modalities like Deep Ultraviolet (DUV) fluorescence. As part of our NIH R01 project, we are pioneering the use of Diffusion Probabilistic Models (DPM) to synthesize high-fidelity, diverse medical images.

Methodology: We employ a two-step diffusion process (forward noise addition and reverse denoising) to generate synthetic DUV patch images that mimic real biological features like cellular density and infiltration.
Impact: Augmenting training data with DPM-generated images increased breast cancer detection accuracy from 93% to 97%, significantly outperforming traditional GAN-based augmentation.

Page updated

Google Sites

Report abuse