Selected work

Representative Papers

SWITi: Quantifying and Reducing Tiling Artifacts with Sliding Window Inner Tiling

Jul 21, 2026

This work addresses the issue of stitching artifacts—often mistaken for genuine structures—in large-image tile-based prediction, which arises from limited receptive fields and independent posterior sampling. The proposed method, SWITi, mitigates these artifacts during inference by averaging predictions over overlapping regions via a sliding window, thereby distributing discrepancies between adjacent tiles across varying locations and preventing artifact accumulation at fixed boundaries, all without requiring additional forward passes. The study introduces, for the first time, no-reference artifact evaluation metrics—FRT and ASV—and integrates pixel-gradient permutation testing to enable automatic detection and quantification of artifacts. Evaluated on both 2D and 3D fluorescence microscopy images, SWITi substantially suppresses stitching seams, enhances reconstruction fidelity and resolution, and effectively supports downstream biomedical analysis.

0 citationsRead paper

Applications of temporal graph learning for predicting the dynamics of biological systems

May 27, 2026

This work addresses the limitation of existing biological foundation models, which predominantly rely on static gene expression data and thus fail to capture the dynamic evolution of gene regulatory networks (GRNs) during cellular development. The authors propose a novel approach that infers pseudotemporal trajectories from single-cell transcriptomic data, discretizes them into developmental snapshots, and reconstructs GRNs at each snapshot. A temporal graph neural network is then introduced to explicitly model the dynamic rewiring of regulatory interactions over time, enabling accurate prediction of gene expression, regulatory links, and key hub genes. To the best of our knowledge, this is the first application of temporal graph learning to single-cell biology. Evaluated on mouse erythroid gastrulation and pancreatic endocrine development datasets, the method significantly outperforms state-of-the-art foundation models such as scGPT and scFoundation across all three tasks, revealing non-trivial dynamic regulatory mechanisms.

0 citationsRead paper

MamaDino: A Hybrid Vision Model for Breast Cancer 3-Year Risk Prediction

Feb 14, 2026

This work proposes MamaDino, a novel approach for breast cancer risk prediction that addresses the limitations of existing models, which rely on high-resolution mammograms and fail to explicitly model bilateral breast asymmetry—leading to performance degradation at lower resolutions. MamaDino uniquely integrates a self-supervised DINOv2 Vision Transformer with a trainable CNN encoder and introduces a BilateralMixer module to explicitly capture asymmetry between left and right breasts. Operating effectively at a reduced resolution of 512×512 (approximately 13× fewer pixels than standard inputs), the method leverages the complementary inductive biases of convolutional and transformer architectures. Evaluated on both internal and external test sets, MamaDino achieves AUCs up to 0.736, matching the performance of the current state-of-the-art model Mirai, while demonstrating robustness across diverse populations and imaging devices.

0 citationsRead paper

indiSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy

Mar 29, 2025

In fluorescence microscopy, the mixing intensity ratios of multi-structure fluorescent signals are unknown and highly variable, severely limiting the generalizability of existing image decomposition methods trained on fixed ratios. To address this, we propose the first mixture-intensity-aware image decomposition framework. Our method introduces three key innovations: (1) a degradation-level regression network that dynamically estimates the severity of signal mixing; (2) a degradation-specific normalization module enabling intensity-adaptive feature calibration; and (3) an end-to-end differentiable architecture based on iterative InDI (Intensity-Dependent Iteration) reconstruction, jointly optimizing decomposition and degradation modeling. Evaluated on five public datasets, our approach unifies solutions for both image splitting and crosstalk removal, consistently outperforming state-of-the-art fixed-ratio methods. It achieves, for the first time, robust decomposition across arbitrary mixing ratios—demonstrating unprecedented generalization and practical applicability in real-world fluorescence imaging scenarios.

0 citationsRead paper

Recent publications

Latest Papers

SWITi: Quantifying and Reducing Tiling Artifacts with Sliding Window Inner Tiling

Jul 21, 2026

This work addresses the issue of stitching artifacts—often mistaken for genuine structures—in large-image tile-based prediction, which arises from limited receptive fields and independent posterior sampling. The proposed method, SWITi, mitigates these artifacts during inference by averaging predictions over overlapping regions via a sliding window, thereby distributing discrepancies between adjacent tiles across varying locations and preventing artifact accumulation at fixed boundaries, all without requiring additional forward passes. The study introduces, for the first time, no-reference artifact evaluation metrics—FRT and ASV—and integrates pixel-gradient permutation testing to enable automatic detection and quantification of artifacts. Evaluated on both 2D and 3D fluorescence microscopy images, SWITi substantially suppresses stitching seams, enhances reconstruction fidelity and resolution, and effectively supports downstream biomedical analysis.

0 citationsRead paper

Applications of temporal graph learning for predicting the dynamics of biological systems

May 27, 2026

This work addresses the limitation of existing biological foundation models, which predominantly rely on static gene expression data and thus fail to capture the dynamic evolution of gene regulatory networks (GRNs) during cellular development. The authors propose a novel approach that infers pseudotemporal trajectories from single-cell transcriptomic data, discretizes them into developmental snapshots, and reconstructs GRNs at each snapshot. A temporal graph neural network is then introduced to explicitly model the dynamic rewiring of regulatory interactions over time, enabling accurate prediction of gene expression, regulatory links, and key hub genes. To the best of our knowledge, this is the first application of temporal graph learning to single-cell biology. Evaluated on mouse erythroid gastrulation and pancreatic endocrine development datasets, the method significantly outperforms state-of-the-art foundation models such as scGPT and scFoundation across all three tasks, revealing non-trivial dynamic regulatory mechanisms.

0 citationsRead paper

MamaDino: A Hybrid Vision Model for Breast Cancer 3-Year Risk Prediction

Feb 14, 2026

This work proposes MamaDino, a novel approach for breast cancer risk prediction that addresses the limitations of existing models, which rely on high-resolution mammograms and fail to explicitly model bilateral breast asymmetry—leading to performance degradation at lower resolutions. MamaDino uniquely integrates a self-supervised DINOv2 Vision Transformer with a trainable CNN encoder and introduces a BilateralMixer module to explicitly capture asymmetry between left and right breasts. Operating effectively at a reduced resolution of 512×512 (approximately 13× fewer pixels than standard inputs), the method leverages the complementary inductive biases of convolutional and transformer architectures. Evaluated on both internal and external test sets, MamaDino achieves AUCs up to 0.736, matching the performance of the current state-of-the-art model Mirai, while demonstrating robustness across diverse populations and imaging devices.

0 citationsRead paper

indiSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy

Mar 29, 2025

In fluorescence microscopy, the mixing intensity ratios of multi-structure fluorescent signals are unknown and highly variable, severely limiting the generalizability of existing image decomposition methods trained on fixed ratios. To address this, we propose the first mixture-intensity-aware image decomposition framework. Our method introduces three key innovations: (1) a degradation-level regression network that dynamically estimates the severity of signal mixing; (2) a degradation-specific normalization module enabling intensity-adaptive feature calibration; and (3) an end-to-end differentiable architecture based on iterative InDI (Intensity-Dependent Iteration) reconstruction, jointly optimizing decomposition and degradation modeling. Evaluated on five public datasets, our approach unifies solutions for both image splitting and crosstalk removal, consistently outperforming state-of-the-art fixed-ratio methods. It achieves, for the first time, robust decomposition across arbitrary mixing ratios—demonstrating unprecedented generalization and practical applicability in real-world fluorescence imaging scenarios.

0 citationsRead paper