For the most up-to-date list, please see my Google Scholar profile
Computational Linguistics Journal, 2025
A psychology-based curriculum learning framework that applies Item Response Theory to artificial crowds for model-independent difficulty quantification. Demonstrated effectiveness on both classification and generation tasks, achieving 69.68% faster training and 75.48% speedup over state-of-the-art methods on large language model fine-tuning.
IEEE Transactions on Medical Imaging (TMI), 2025
Ceb, a boundary-aware approach to cell instance segmentation that clusters foreground pixels by classifying cell boundaries instead of relying on pixel-wise objectives, which often discard geometric properties such as shape, curvature, and convexity. Starting from semantic segmentation probability maps, Ceb generates candidate boundaries with a revised Watershed algorithm, encodes each as a novel "boundary signature," and trains a lightweight CNN to separate true from false boundaries; a temporal-consistency extension further improves performance on cell video datasets. Across six public datasets, Ceb consistently outperforms existing foreground pixel-clustering methods and is competitive with or better than state-of-the-art cell instance segmentation methods.
IEEE Transactions on Multimedia (TMM), 2025
A GPU-accelerated Nearest Neighbor Search approximation algorithm achieving 44× to 135× speedup over exact Earth Mover's Distance computation while maintaining superior accuracy (91.88% on 20news, 93.24% on Amazon reviews). Includes comprehensive theoretical analysis of time complexity and error bounds, enabling scalable optimal transport for large-scale computer vision applications.
European Conference on Computer Vision (ECCV), 2026
The first LLM-based agentic framework for automated topological descriptor selection in medical imaging. Operating through a Perception-Reasoning-Action-Reflection loop with 21 domain-specific tools and dual memory, TopoAgent selects from 15 topological descriptors, configures parameters, and returns features with a full reasoning trace. On TopoBenchmark (113,182 samples, 5 morphological categories), it achieves 68.21% balanced accuracy, outperforming the strongest baseline by 9.32% and general-purpose LLMs by over 21%.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
A universal framework that demonstrates consistent accuracy improvements averaging +3.26% with strong statistical significance (80% of comparisons with p<0.001) across five state-of-the-art contrastive learning methods (SimCLR, MoCo-v3, BYOL, DINO, and Barlow Twins) on five medical imaging datasets. Features topology-aware augmentations, a Hierarchical Topology Encoder, and an adaptive mixture-of-experts fusion module.
SIAM International Conference on Data Mining (SDM), 2022
Novel graph-based approach for predicting traffic accident impact using single snapshot observations, leveraging abrupt signal modeling in graph neural networks.
Under review at WACV 2027
A novel visual-structural and topology-aware framework that achieves state-of-the-art performance with 81.51% average accuracy across 13 medical benchmarks, representing a 4.48% improvement over existing baselines while nearly matching supervised methods in a fully unsupervised setting. The framework adapts SAM with mask-based structural priors and persistent homology features to enable anatomically consistent clustering.
Under review at Journal of Applied and Computational Topology
A stable algorithmic framework for comparing Reeb graphs using Gromov-Wasserstein distance with persistence images, accompanied by rigorous theoretical stability proofs under scalar field perturbations. This work provides theoretical guarantees for topological shape analysis and comparison in computational geometry applications.
Under review at IEEE Transactions on Visualization and Computer Graphics (TVCG)
A framework for comparing Morse-Smale complexes by representing them as attributed hypergraphs and applying co-optimal transport with a type-aware penalty to jointly align critical points and their connectivity. Evaluated on scientific visualization datasets including Vortex Street, Ionization Front, Heated Cylinder, and Viscous Finger, enabling structure-aware comparison of scalar field topology.