Selected work

Representative Papers

Benchmarking and Evaluation of AI Models in Biology: Outcomes and Recommendations from the CZI Virtual Cells Workshop

Jul 14, 2025

Biology lacks cross-domain, standardized AI model benchmarks, hindering model robustness and trustworthiness. To address this, we introduce the first multimodal AI benchmarking framework spanning imaging, transcriptomics, proteomics, and genomics—systematically tackling data heterogeneity, noise, bias, and resource fragmentation. Our approach integrates a high-fidelity data curation pipeline, unified preprocessing tools, biologically grounded multimodal evaluation metrics, and an open collaborative platform to enable fair, cross-task and cross-modal comparisons. A core innovation is the “virtual cell” paradigm—a biologically anchored, integrative evaluation framework—that unifies disparate modalities through shared cellular context. We further release a reproducible, extensible set of AI model evaluation guidelines. The framework significantly enhances rigor, transparency, and cross-domain comparability in biological AI research, accelerating AI-driven mechanistic discovery and therapeutic translation.

0 citationsRead paper

Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels

May 28, 2025

CT reconstruction kernel differences introduce systematic bias in pulmonary quantitative analysis—e.g., emphysema scoring—particularly under low-dose scanning, compromising clinical comparability across scanners and protocols. To address this, we propose a multi-path cycleGAN framework with a shared latent space, the first to jointly model both paired and unpaired cross-kernel CT data translation. Our method incorporates kernel-specific encoders/decoders, a multi-discriminator architecture, and anatomical fidelity constraints guided by TotalSegmentator segmentation. Experiments demonstrate statistically significant reduction in emphysema score variability (p < 0.05) and elimination of confounding inter-kernel differences in unpaired settings (p > 0.05). Muscle and fat segmentation achieves Dice coefficients > 0.9; vascular structure overlap remains anatomically plausible. The approach outperforms conventional and switchable cycleGAN baselines in both quantitative accuracy and anatomical consistency.

0 citationsRead paper

Investigating the impact of kernel harmonization and deformable registration on inspiratory and expiratory chest CT images for people with COPD

Feb 07, 2025

This study addresses reconstruction kernel-induced bias in emphysema quantification during inspiratory–expiratory CT registration in COPD patients, arising from mismatched kernels (e.g., BONE vs. STANDARD). We propose a two-stage “kernel harmonization + deformable registration” framework: first, CycleGAN is employed to translate hard-kernel inspiratory images into soft-kernel style; second, voxel-based deformable registration is performed on harmonized images. To our knowledge, this is the first work jointly modeling kernel harmonization and registration to enhance quantitative robustness. Evaluated on the COPDGene dataset, kernel harmonization reduced the median emphysema fraction from 10.479% to 3.039%, approaching the reference soft-kernel value (1.305%). Subsequent registration significantly improved Dice similarity coefficients of emphysema masks (p < 0.001), demonstrating effective suppression of kernel-related artifacts and enhanced accuracy in regional volume change analysis.

0 citationsRead paper

Fast Vision Mamba: Pooling Spatial Dimensions for Accelerated Processing

Feb 01, 2025

Vision Mamba and other vision large models suffer from low inference efficiency and sequence-length limitations when processing ultra-high-resolution images (e.g., 2048×2048). Method: This paper proposes FastVim, a novel architecture that introduces cross-Mamba-block alternating spatial-dimension token pooling—a first-of-its-kind dynamic downsampling strategy—halving the parallel scanning steps of state space models (SSMs) and overcoming the linear sequence modeling bottleneck. FastVim integrates SSMs with selective scanning to enable spatially aware computational compression. Contribution/Results: Extensive experiments demonstrate that FastVim achieves state-of-the-art accuracy across diverse vision tasks—including image classification, semantic segmentation, object detection, and cellular perturbation prediction—while delivering a 72.5% speedup in inference latency. This significantly improves throughput and scalability for ultra-high-resolution vision workloads.

0 citationsRead paper

Robust Body Composition Analysis by Generating 3D CT Volumes from Limited 2D Slices

Jan 22, 2025

To address the instability of body composition analysis arising from missing spatial information in single-slice 2D CT images, this paper proposes a sparse-slice 3D reconstruction method that synergistically integrates a latent diffusion model (LDM) guided by anatomical location regression and a variational autoencoder (VAE), enabling generation of physically plausible and high-fidelity 3D CT volumes from a minimal number of 2D slices. Unlike conventional interpolation or single-slice analysis, our approach is the first to embed anatomical priors directly into the diffusion process to ensure anatomical consistency of generated volumes. Evaluated on multicenter data, the method reduces mean body composition analysis error from 23.3% to 15.2%, while substantially improving robustness and reproducibility in fat and muscle quantification. This work establishes a novel paradigm for accurate body composition assessment from low-dose CT scans.

0 citationsRead paper

Recent publications

Latest Papers

Benchmarking and Evaluation of AI Models in Biology: Outcomes and Recommendations from the CZI Virtual Cells Workshop

Jul 14, 2025

Biology lacks cross-domain, standardized AI model benchmarks, hindering model robustness and trustworthiness. To address this, we introduce the first multimodal AI benchmarking framework spanning imaging, transcriptomics, proteomics, and genomics—systematically tackling data heterogeneity, noise, bias, and resource fragmentation. Our approach integrates a high-fidelity data curation pipeline, unified preprocessing tools, biologically grounded multimodal evaluation metrics, and an open collaborative platform to enable fair, cross-task and cross-modal comparisons. A core innovation is the “virtual cell” paradigm—a biologically anchored, integrative evaluation framework—that unifies disparate modalities through shared cellular context. We further release a reproducible, extensible set of AI model evaluation guidelines. The framework significantly enhances rigor, transparency, and cross-domain comparability in biological AI research, accelerating AI-driven mechanistic discovery and therapeutic translation.

0 citationsRead paper

Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels

May 28, 2025

CT reconstruction kernel differences introduce systematic bias in pulmonary quantitative analysis—e.g., emphysema scoring—particularly under low-dose scanning, compromising clinical comparability across scanners and protocols. To address this, we propose a multi-path cycleGAN framework with a shared latent space, the first to jointly model both paired and unpaired cross-kernel CT data translation. Our method incorporates kernel-specific encoders/decoders, a multi-discriminator architecture, and anatomical fidelity constraints guided by TotalSegmentator segmentation. Experiments demonstrate statistically significant reduction in emphysema score variability (p < 0.05) and elimination of confounding inter-kernel differences in unpaired settings (p > 0.05). Muscle and fat segmentation achieves Dice coefficients > 0.9; vascular structure overlap remains anatomically plausible. The approach outperforms conventional and switchable cycleGAN baselines in both quantitative accuracy and anatomical consistency.

0 citationsRead paper

Investigating the impact of kernel harmonization and deformable registration on inspiratory and expiratory chest CT images for people with COPD

Feb 07, 2025

This study addresses reconstruction kernel-induced bias in emphysema quantification during inspiratory–expiratory CT registration in COPD patients, arising from mismatched kernels (e.g., BONE vs. STANDARD). We propose a two-stage “kernel harmonization + deformable registration” framework: first, CycleGAN is employed to translate hard-kernel inspiratory images into soft-kernel style; second, voxel-based deformable registration is performed on harmonized images. To our knowledge, this is the first work jointly modeling kernel harmonization and registration to enhance quantitative robustness. Evaluated on the COPDGene dataset, kernel harmonization reduced the median emphysema fraction from 10.479% to 3.039%, approaching the reference soft-kernel value (1.305%). Subsequent registration significantly improved Dice similarity coefficients of emphysema masks (p < 0.001), demonstrating effective suppression of kernel-related artifacts and enhanced accuracy in regional volume change analysis.

0 citationsRead paper

Fast Vision Mamba: Pooling Spatial Dimensions for Accelerated Processing

Feb 01, 2025

Vision Mamba and other vision large models suffer from low inference efficiency and sequence-length limitations when processing ultra-high-resolution images (e.g., 2048×2048). Method: This paper proposes FastVim, a novel architecture that introduces cross-Mamba-block alternating spatial-dimension token pooling—a first-of-its-kind dynamic downsampling strategy—halving the parallel scanning steps of state space models (SSMs) and overcoming the linear sequence modeling bottleneck. FastVim integrates SSMs with selective scanning to enable spatially aware computational compression. Contribution/Results: Extensive experiments demonstrate that FastVim achieves state-of-the-art accuracy across diverse vision tasks—including image classification, semantic segmentation, object detection, and cellular perturbation prediction—while delivering a 72.5% speedup in inference latency. This significantly improves throughput and scalability for ultra-high-resolution vision workloads.

0 citationsRead paper

Robust Body Composition Analysis by Generating 3D CT Volumes from Limited 2D Slices

Jan 22, 2025

To address the instability of body composition analysis arising from missing spatial information in single-slice 2D CT images, this paper proposes a sparse-slice 3D reconstruction method that synergistically integrates a latent diffusion model (LDM) guided by anatomical location regression and a variational autoencoder (VAE), enabling generation of physically plausible and high-fidelity 3D CT volumes from a minimal number of 2D slices. Unlike conventional interpolation or single-slice analysis, our approach is the first to embed anatomical priors directly into the diffusion process to ensure anatomical consistency of generated volumes. Evaluated on multicenter data, the method reduces mean body composition analysis error from 23.3% to 15.2%, while substantially improving robustness and reproducibility in fat and muscle quantification. This work establishes a novel paradigm for accurate body composition assessment from low-dose CT scans.

0 citationsRead paper