Investigating the geometric and topological structure of latent manifolds learned from real and synthetic medical images to better understand representation quality and generative behavior. The project combines manifold learning, dimensionality reduction, and visualization-driven analysis to improve interpretability, realism assessment, and controllable medical image generation.
Developed an anatomy-aware and pathology-controlled 3D diffusion framework for lung CT synthesis, enabling realistic generation of clinically consistent pulmonary lesions with precise anatomical control. The project combines mask-conditioned diffusion modeling with topology-aware structural guidance to improve lesion fidelity, anatomical realism, and trustworthiness in AI-driven medical imaging.
Analyzed the loss landscape of neural network training using topological data analysis techniques such as contour trees, join trees, and split tree simplification to interpret training dynamics.
Developed an unsupervised segmentation method using eigenvector variation across intrinsic dimensions to distinguish intersecting manifolds in high-dimensional data.
View Project
Segmented and localized lumbar spine tumors from T2-weighted MRI scans using deep learning-based medical image analysis techniques.
View Project
Designed a novel lightweight feature extraction model to detect plant diseases from leaf images, optimized for low-computation environments. Presented at ICSTA 2023.
View Project