Selected work

Representative Papers

Integrating Implicit and Explicit Relational Biases through Graph-Based Multiple Instance Learning: A Case Study in Skin Lesion Diagnosis

Aug 06, 2026

This work addresses the insufficient modeling of structural relationships in skin lesion image classification by proposing a dual-level relational fusion framework. At the implicit level, a convolutional masked autoencoder is employed to learn self-supervised inter-patch relationships, while at the explicit level, a graph attention network leverages multiple graph topologies—such as grid and k-nearest neighbor structures—to facilitate message passing. This approach represents the first effort to jointly exploit implicit self-supervised relations and explicit graph-structured priors, thereby enhancing the model’s structural awareness. The method achieves balanced accuracies of 79.27% and 60.67% on the ISIC-2018 and ISIC-2019 datasets, respectively, significantly outperforming both single-relation modeling strategies and baseline approaches.

0 citationsRead paper

CARDIAG: A Dense Segment Classification Benchmark of Deep Learning Architectures for Coronary Angiography

Jul 24, 2026

This study addresses the absence of a standardized benchmark for pixel-level SYNTAX category segmentation in coronary angiography, which has hindered fair comparison of deep learning models. To this end, we introduce CARDIAG, a multicenter, multilabel dense classification benchmark that establishes the first standardized evaluation framework tailored for SYNTAX scoring. It encompasses 24 state-of-the-art architectures—including ConvNeXt V2, Mamba U-Net, DeepLabV3+, and FPN—and provides a high-quality dataset annotated with SYNTAX labels, segmentation masks, uncertainty estimates, and DICOM metadata. Through rigorous data partitioning and multidimensional metrics such as diameter error and centerline quality, we systematically assess model calibration, generalization, and data efficiency. The best single model (ConvNeXt V2 + DeepLabV3+) achieves a macro F₁ score of 0.456, improving to 0.479 with ensembling, while all models demonstrate strong calibration and robust cross-center performance.

0 citationsRead paper

Hierarchical $\mathcal{F}$-Clustering: Approximation and Hardness of Clustering into Trees and Bounded Diameter Graphs

Jul 14, 2026

This work addresses the hierarchical $\mathcal{F}$-clustering problem, where the stopping condition is relaxed from singleton clusters to clusters belonging to a specified graph class—such as trees or graphs of bounded diameter. The authors propose a unified linear programming–based approximation framework that models the underlying flat clustering as an integer linear program and employs tailored rounding strategies to construct a clustering tree. This framework yields the first $\mathcal{O}(\log n \cdot \log \log n)$-approximation algorithm for tree clustering and an $\mathcal{O}(\log n)$-approximation for bounded-diameter graph clustering, while also characterizing structural conditions on graph classes under which the approach applies. Furthermore, under the Small Set Expansion hypothesis, the paper establishes hardness results showing that constant-factor approximations are infeasible, thereby delineating tight theoretical bounds.

0 citationsRead paper

Graph Partitioning with Demands: Generalized Conductance and its Applications

Jul 14, 2026

This work introduces and systematically studies the generalized conductance minimization problem: given a graph with vertex demands and edge capacities, find a cut that minimizes generalized conductance while ensuring the total demand in each connected component does not exceed a prescribed upper bound. The authors design a combinatorial approximation algorithm via a two-stage reduction—first to a generalized $k$-multicut problem, then to a sparsest cut problem with demand constraints. Their main contributions include establishing novel connections between generalized conductance and classical graph partitioning problems, extending the framework to demand-aware graph partitioning and hierarchical clustering, obtaining an $O(\log n)$-approximation algorithm, improving this to $O(\sqrt{\log n})$ for multiplicative demand functions, and achieving a constant-factor approximation on trees.

0 citationsRead paper

NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices

Jul 11, 2026

This work addresses the high computational cost of existing video super-resolution methods, which typically rely on Transformers or explicit optical flow estimation, hindering deployment on edge devices. To overcome this limitation, the authors propose NanoVSR, a lightweight fully convolutional architecture that leverages structural reparameterization to employ only standard convolutions during inference. Combined with a progressive training strategy, NanoVSR implicitly learns spatiotemporal alignment without explicit motion compensation, achieving high-quality super-resolution while maintaining efficiency and hardware acceleration compatibility. Experimental results demonstrate that NanoVSR-644k attains 28.64 dB PSNR on REDS4 and achieves real-time performance (27.2 FPS) on a Jetson Orin NX; the larger NanoVSR-1.7M variant further improves PSNR to 29.15 dB while sustaining 19.58 FPS.

0 citationsRead paper

Recent publications

Latest Papers

Integrating Implicit and Explicit Relational Biases through Graph-Based Multiple Instance Learning: A Case Study in Skin Lesion Diagnosis

Aug 06, 2026

This work addresses the insufficient modeling of structural relationships in skin lesion image classification by proposing a dual-level relational fusion framework. At the implicit level, a convolutional masked autoencoder is employed to learn self-supervised inter-patch relationships, while at the explicit level, a graph attention network leverages multiple graph topologies—such as grid and k-nearest neighbor structures—to facilitate message passing. This approach represents the first effort to jointly exploit implicit self-supervised relations and explicit graph-structured priors, thereby enhancing the model’s structural awareness. The method achieves balanced accuracies of 79.27% and 60.67% on the ISIC-2018 and ISIC-2019 datasets, respectively, significantly outperforming both single-relation modeling strategies and baseline approaches.

0 citationsRead paper

CARDIAG: A Dense Segment Classification Benchmark of Deep Learning Architectures for Coronary Angiography

Jul 24, 2026

This study addresses the absence of a standardized benchmark for pixel-level SYNTAX category segmentation in coronary angiography, which has hindered fair comparison of deep learning models. To this end, we introduce CARDIAG, a multicenter, multilabel dense classification benchmark that establishes the first standardized evaluation framework tailored for SYNTAX scoring. It encompasses 24 state-of-the-art architectures—including ConvNeXt V2, Mamba U-Net, DeepLabV3+, and FPN—and provides a high-quality dataset annotated with SYNTAX labels, segmentation masks, uncertainty estimates, and DICOM metadata. Through rigorous data partitioning and multidimensional metrics such as diameter error and centerline quality, we systematically assess model calibration, generalization, and data efficiency. The best single model (ConvNeXt V2 + DeepLabV3+) achieves a macro F₁ score of 0.456, improving to 0.479 with ensembling, while all models demonstrate strong calibration and robust cross-center performance.

0 citationsRead paper

Hierarchical $\mathcal{F}$-Clustering: Approximation and Hardness of Clustering into Trees and Bounded Diameter Graphs

Jul 14, 2026

This work addresses the hierarchical $\mathcal{F}$-clustering problem, where the stopping condition is relaxed from singleton clusters to clusters belonging to a specified graph class—such as trees or graphs of bounded diameter. The authors propose a unified linear programming–based approximation framework that models the underlying flat clustering as an integer linear program and employs tailored rounding strategies to construct a clustering tree. This framework yields the first $\mathcal{O}(\log n \cdot \log \log n)$-approximation algorithm for tree clustering and an $\mathcal{O}(\log n)$-approximation for bounded-diameter graph clustering, while also characterizing structural conditions on graph classes under which the approach applies. Furthermore, under the Small Set Expansion hypothesis, the paper establishes hardness results showing that constant-factor approximations are infeasible, thereby delineating tight theoretical bounds.

0 citationsRead paper

Graph Partitioning with Demands: Generalized Conductance and its Applications

Jul 14, 2026

This work introduces and systematically studies the generalized conductance minimization problem: given a graph with vertex demands and edge capacities, find a cut that minimizes generalized conductance while ensuring the total demand in each connected component does not exceed a prescribed upper bound. The authors design a combinatorial approximation algorithm via a two-stage reduction—first to a generalized $k$-multicut problem, then to a sparsest cut problem with demand constraints. Their main contributions include establishing novel connections between generalized conductance and classical graph partitioning problems, extending the framework to demand-aware graph partitioning and hierarchical clustering, obtaining an $O(\log n)$-approximation algorithm, improving this to $O(\sqrt{\log n})$ for multiplicative demand functions, and achieving a constant-factor approximation on trees.

0 citationsRead paper

NanoVSR: Towards Real-Time Video Super-Resolution on Edge Devices

Jul 11, 2026

This work addresses the high computational cost of existing video super-resolution methods, which typically rely on Transformers or explicit optical flow estimation, hindering deployment on edge devices. To overcome this limitation, the authors propose NanoVSR, a lightweight fully convolutional architecture that leverages structural reparameterization to employ only standard convolutions during inference. Combined with a progressive training strategy, NanoVSR implicitly learns spatiotemporal alignment without explicit motion compensation, achieving high-quality super-resolution while maintaining efficiency and hardware acceleration compatibility. Experimental results demonstrate that NanoVSR-644k attains 28.64 dB PSNR on REDS4 and achieves real-time performance (27.2 FPS) on a Jetson Orin NX; the larger NanoVSR-1.7M variant further improves PSNR to 29.15 dB while sustaining 19.58 FPS.

0 citationsRead paper