Chengze Ye
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Chengze Ye has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Yipeng Sun
Friedrich-Alexander-Universität Erlangen-Nürnberg, Pattern Recognition Lab, Erlangen, Germany
10 shared publications - Andreas Maier
Friedrich-Alexander-Universität Erlangen-Nürnberg, Pattern Recognition Lab, Erlangen, Germany
9 shared publications - Linda-Sophie Schneider
Friedrich-Alexander-Universität Erlangen-Nürnberg, Pattern Recognition Lab, Erlangen, Germany
9 shared publications - Siyuan Mei
Pattern Recognition Lab, FAU Erlangen-Nürnberg
7 shared publications - Siming Bayer
Friedrich-Alexander-Universität Erlangen-Nürnberg, Pattern Recognition Lab, Erlangen, Germany
6 shared publications Andreas Maier
4 shared publications
Latest Publications
- 1 November 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (06) , 1-23
Noise2Flow: single-image conditional mean denoising flows from weak-anchor couplings
PurposeSingle-image denoising is needed when clean targets and repeated acquisitions are unavailable, as in low-dose computed tomography (CT) and fluorescence microscopy. Existing per-image methods often assume pixel-wise noise independence, which fails under spatially correlated noise, and imperfect single-image denoisers commonly leave residual noise. We learn a proxy mean-velocity direction field from one noisy image so that a single transport step exceeds the imperfect anchors used as supervision.ApproachNoise2Flow runs weak single-image denoisers on the noisy image to form source–target couplings. The displacement from input to anchor is a directional proxy for the mean restoration velocity. For each imposed linear segment, MeanFlow’s finite-interval average velocity reduces exactly to a constant secant. A flow network is trained with a direction-normalized objective that regularizes predictions at states in a stochastic tube around the noisy-to-anchor segment toward a common secant direction. At inference, the weak denoisers are discarded; one update with a calibrated transport budget restores the image.ResultsOn the fluorescence microscopy denoising (FMD) test set of 48 images, Noise2Flow reaches 34.91 dB peak signal-to-noise ratio and 0.887 structural similarity, 0.52 dB above Filter2Noise, and 1.00 dB above Noise2Detail. On 1311 low-dose CT slices from 10 Mayo Clinic patients, it reaches 28.82 dB and 0.714, 0.91 dB above Filter2Noise and 1.23 dB above Noise2Detail. One-step inference is within 0.11 dB of 10-step Heun; tube sampling improves the 10-step Heun result relative to on-segment training; a calibrated extrapolation factor moves the flow past residual-noise anchors of varying quality.ConclusionsA proxy mean-velocity direction field can be learned from a single noisy medical image by treating weak-denoiser outputs as directional anchors rather than final targets. The straight proxy path defines the one-step endpoint update; tube regularization and calibrated extrapolation carry that update past residual-noise anchors. The learned flow exceeds residual-noise anchors and supports reference-free restoration in data-limited imaging.
- 1 November 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (06) , 1-23
Noise2Flow: single-image conditional mean denoising flows from weak-anchor couplings
PurposeSingle-image denoising is needed when clean targets and repeated acquisitions are unavailable, as in low-dose computed tomography (CT) and fluorescence microscopy. Existing per-image methods often assume pixel-wise noise independence, which fails under spatially correlated noise, and imperfect single-image denoisers commonly leave residual noise. We learn a proxy mean-velocity direction field from one noisy image so that a single transport step exceeds the imperfect anchors used as supervision.ApproachNoise2Flow runs weak single-image denoisers on the noisy image to form source–target couplings. The displacement from input to anchor is a directional proxy for the mean restoration velocity. For each imposed linear segment, MeanFlow’s finite-interval average velocity reduces exactly to a constant secant. A flow network is trained with a direction-normalized objective that regularizes predictions at states in a stochastic tube around the noisy-to-anchor segment toward a common secant direction. At inference, the weak denoisers are discarded; one update with a calibrated transport budget restores the image.ResultsOn the fluorescence microscopy denoising (FMD) test set of 48 images, Noise2Flow reaches 34.91 dB peak signal-to-noise ratio and 0.887 structural similarity, 0.52 dB above Filter2Noise, and 1.00 dB above Noise2Detail. On 1311 low-dose CT slices from 10 Mayo Clinic patients, it reaches 28.82 dB and 0.714, 0.91 dB above Filter2Noise and 1.23 dB above Noise2Detail. One-step inference is within 0.11 dB of 10-step Heun; tube sampling improves the 10-step Heun result relative to on-segment training; a calibrated extrapolation factor moves the flow past residual-noise anchors of varying quality.ConclusionsA proxy mean-velocity direction field can be learned from a single noisy medical image by treating weak-denoiser outputs as directional anchors rather than final targets. The straight proxy path defines the one-step endpoint update; tube regularization and calibrated extrapolation carry that update past residual-noise anchors. The learned flow exceeds residual-noise anchors and supports reference-free restoration in data-limited imaging.
- Open Access
- 14 October 2026
- Published by MDPI AG in Preprints.org
diffct: Differentiable CT Operators from Circular Orbits to Arbitrary Trajectories
Diffct implements computed-tomography projection and reconstruction in PyTorch using Numba CUDA kernels. Its native forward model integrates a cell-constant attenuation field along rays defined by per-view acquisition geometry; the backprojector is the matched discrete adjoint. The same model supports parallel, fan and cone beams, non-circular trajectories, and parameterized detector surfaces. This report describes the implementation through release 2.2.1: the geometry and sampling conventions, Siddon traversal, data and first-order geometry derivatives, differentiable analytical FBP/FDK, and execution on one or several GPUs. We derive the ownership and reduction rules for distributed view and spatial partitions and explain how spatial, view and detector blocks extend native operations to CPU-backed and disk-backed data. Experiments include measured walnut cone-beam data, synthetic reconstruction and calibration, and a recorded 512³ scaling benchmark. On A100-SXM-64GB GPUs, eight GPUs on two nodes reduce forward projection from 493.05 to 73.18 ms and the amortized cost of a ten-step CGLS solve from 2054.36 to 373.38 ms per iteration. Analytical gathers have their own discrete transposes; differentiability does not make FDK an exact inverse. We distinguish numerical correctness, reconstruction quality and execution cost, and state the limits of each experiment.
- 14 October 2026
- Published by MDPI AG in Preprints.org
Open Accessdiffct: Differentiable CT Operators from Circular Orbits to Arbitrary Trajectories
Diffct implements computed-tomography projection and reconstruction in PyTorch using Numba CUDA kernels. Its native forward model integrates a cell-constant attenuation field along rays defined by per-view acquisition geometry; the backprojector is the matched discrete adjoint. The same model supports parallel, fan and cone beams, non-circular trajectories, and parameterized detector surfaces. This report describes the implementation through release 2.2.1: the geometry and sampling conventions, Siddon traversal, data and first-order geometry derivatives, differentiable analytical FBP/FDK, and execution on one or several GPUs. We derive the ownership and reduction rules for distributed view and spatial partitions and explain how spatial, view and detector blocks extend native operations to CPU-backed and disk-backed data. Experiments include measured walnut cone-beam data, synthetic reconstruction and calibration, and a recorded 512³ scaling benchmark. On A100-SXM-64GB GPUs, eight GPUs on two nodes reduce forward projection from 493.05 to 73.18 ms and the amortized cost of a ten-step CGLS solve from 2054.36 to 373.38 ms per iteration. Analytical gathers have their own discrete transposes; differentiability does not make FDK an exact inverse. We distinguish numerical correctness, reconstruction quality and execution cost, and state the limits of each experiment.
- 26 September 2026
- Published by Springer Nature
- p. 552-562
TCCT: Trajectory-Conditioned CBCT Reconstruction for Sinusoidal Non-circular Orbits with a Fourier Neural Operator
- 26 September 2026
- Published by Springer Nature
- p. 552-562
TCCT: Trajectory-Conditioned CBCT Reconstruction for Sinusoidal Non-circular Orbits with a Fourier Neural Operator
- Open Access
- 27 April 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Vision transformer Hook for dense predictions
- 27 April 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Open AccessVision transformer Hook for dense predictions
- 12 March 2026
- Published by Springer Nature
- p. 328-333
Differentiable Approximate Truncation Robust CBCT Reconstruction via Known Operator Learning
- 12 March 2026
- Published by Springer Nature
- p. 328-333
Differentiable Approximate Truncation Robust CBCT Reconstruction via Known Operator Learning
- 12 March 2026
- Published by Springer Nature
- p. 34-39
Interpretable Framework for Zero-shot 4D Low-dose CT Denoising
- 12 March 2026
- Published by Springer Nature
- p. 34-39
Interpretable Framework for Zero-shot 4D Low-dose CT Denoising
- Open Access
- 1 March 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (02) , 1-24
Filter2Noise: a framework for interpretable and zero-shot low-dose CT image denoising
PurposeDeep learning has achieved remarkable progress in low-dose computed tomography (LDCT) denoising; however, radiologists struggle to trust black-box models they cannot verify or control. Zero-shot methods eliminate training data requirements but fail on computed tomography’s (CT) spatially correlated noise. We demonstrate that a transparent mathematical operator, when made content-adaptive, can match deep learning performance while remaining fully interpretable.ApproachWe introduce Filter2Noise (F2N), which replaces conventional deep networks with an attention-guided bilateral filter that adapts to local anatomy. A lightweight attention module (3.6k parameters) predicts optimal filtering strategies for each image region by analyzing tissue type, texture, and noise characteristics. To enable robust learning from a single noisy image with correlated noise, we develop Euclidean local shuffle, which strategically disrupts noise correlations while preserving anatomical structure, and a multi-scale self-supervised loss that enforces consistency across resolutions.ResultsOn the Mayo Clinic LDCT Grand Challenge, F2N achieves 39.76 dB peak signal-to-noise ratio, outperforming the next-best zero-shot method by 1.88 dB, while using 360× fewer parameters (3.6k versus 1.3M). Clinical validation on photon-counting CT demonstrates that F2N elevates low-dose images to full-dose quality (no statistically significant difference in contrast-to-noise ratio, p=0.10). The learned filtering strategy is fully visualizable: parameter maps reveal content-aware behavior. Radiologists can interactively adjust these parameters post-training to refine denoising in diagnostically critical regions.ConclusionsF2N reconciles competitive performance with complete interpretability and user control, providing radiologists with a verifiable tool that works across scanners and protocols without retraining.
- 1 March 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 13 (02) , 1-24
Open AccessFilter2Noise: a framework for interpretable and zero-shot low-dose CT image denoising
PurposeDeep learning has achieved remarkable progress in low-dose computed tomography (LDCT) denoising; however, radiologists struggle to trust black-box models they cannot verify or control. Zero-shot methods eliminate training data requirements but fail on computed tomography’s (CT) spatially correlated noise. We demonstrate that a transparent mathematical operator, when made content-adaptive, can match deep learning performance while remaining fully interpretable.ApproachWe introduce Filter2Noise (F2N), which replaces conventional deep networks with an attention-guided bilateral filter that adapts to local anatomy. A lightweight attention module (3.6k parameters) predicts optimal filtering strategies for each image region by analyzing tissue type, texture, and noise characteristics. To enable robust learning from a single noisy image with correlated noise, we develop Euclidean local shuffle, which strategically disrupts noise correlations while preserving anatomical structure, and a multi-scale self-supervised loss that enforces consistency across resolutions.ResultsOn the Mayo Clinic LDCT Grand Challenge, F2N achieves 39.76 dB peak signal-to-noise ratio, outperforming the next-best zero-shot method by 1.88 dB, while using 360× fewer parameters (3.6k versus 1.3M). Clinical validation on photon-counting CT demonstrates that F2N elevates low-dose images to full-dose quality (no statistically significant difference in contrast-to-noise ratio, p=0.10). The learned filtering strategy is fully visualizable: parameter maps reveal content-aware behavior. Radiologists can interactively adjust these parameters post-training to refine denoising in diagnostically critical regions.ConclusionsF2N reconciles competitive performance with complete interpretability and user control, providing radiologists with a verifiable tool that works across scanners and protocols without retraining.
- 1 February 2026
- Published by Machine Learning for Biomedical Imaging in Machine Learning for Biomedical Imaging
- Vol. 2026 (MELBA–BVM ) , 284-296
Robustness and Stability Analysis of Differentiable Shift-Variant FBP for Cone-Beam CT under Challenging Acquisition Settings
The differentiable shift-variant filtered backprojection (SV-FBP) framework enables data-driven estimation of redundancy weights for cone-beam CT reconstruction under general source trajectories, removing the need for analytically derived weighting schemes. In this work, we present a systematic study of the robustness and adaptability of differentiable SV-FBP under challenging acquisition settings. We show that the framework remains stable across highly irregular and discontinuous trajectories, indicating that reconstruction performance is largely insensitive to trajectory ordering or continuity. Instead, the spatial distribution of sampling points plays a more dominant role. Under sparse-view conditions, differentiable SV-FBP achieves competitive reconstruction quality while providing an order-of-magnitude reduction in computation time compared to iterative reconstruction methods at moderate sampling densities. However, we identify a clear transition regime under severe undersampling, where the absence of iterative data consistency leads to performance degradation. Furthermore, we demonstrate that the framework remains applicable to non-planar multi-isocenter geometries, such as Lissajous-saddle trajectories, without requiring architectural modifications. These findings provide new insights into the behavior and limitations of the differentiable SV-FBP model and highlight it as a flexible and efficient solution for non-standard and robotic CBCT acquisition scenarios.
- 1 February 2026
- Published by Machine Learning for Biomedical Imaging in Machine Learning for Biomedical Imaging
- Vol. 2026 (MELBA–BVM ) , 284-296
Robustness and Stability Analysis of Differentiable Shift-Variant FBP for Cone-Beam CT under Challenging Acquisition Settings
The differentiable shift-variant filtered backprojection (SV-FBP) framework enables data-driven estimation of redundancy weights for cone-beam CT reconstruction under general source trajectories, removing the need for analytically derived weighting schemes. In this work, we present a systematic study of the robustness and adaptability of differentiable SV-FBP under challenging acquisition settings. We show that the framework remains stable across highly irregular and discontinuous trajectories, indicating that reconstruction performance is largely insensitive to trajectory ordering or continuity. Instead, the spatial distribution of sampling points plays a more dominant role. Under sparse-view conditions, differentiable SV-FBP achieves competitive reconstruction quality while providing an order-of-magnitude reduction in computation time compared to iterative reconstruction methods at moderate sampling densities. However, we identify a clear transition regime under severe undersampling, where the absence of iterative data consistency leads to performance degradation. Furthermore, we demonstrate that the framework remains applicable to non-planar multi-isocenter geometries, such as Lissajous-saddle trajectories, without requiring architectural modifications. These findings provide new insights into the behavior and limitations of the differentiable SV-FBP model and highlight it as a flexible and efficient solution for non-standard and robotic CBCT acquisition scenarios.
- Open Access
- 20 March 2025
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 70 (7) , 075005
DRACO: differentiable reconstruction for arbitrary CBCT orbits
Objective. This study introduces a novel method for reconstructing cone beam computed tomography (CBCT) images for arbitrary orbits, addressing the computational and memory challenges associated with traditional iterative reconstruction algorithms. Approach. The proposed method employs a differentiable shift-variant filtered backprojection neural network, optimized for arbitrary trajectories. By integrating known operators into the learning model, the approach minimizes the number of trainable parameters while enhancing model interpretability. This framework adapts seamlessly to specific orbit geometries, including non-continuous trajectories such as circular-plus-arc or sinusoidal paths, enabling faster and more accurate CBCT reconstructions. Main results. Experimental validation demonstrates that the method significantly accelerates reconstruction, reducing computation time by over 97% compared to conventional iterative algorithms. It achieves superior or comparable image quality with reduced noise, as evidenced by a 38.6% reduction in mean squared error, a 7.7% increase in peak signal-to-noise ratio, and a 5.0% improvement in the structural similarity index measure. The flexibility and robustness of the approach are confirmed through its ability to handle data from diverse scan geometries. Significance. This method represents a significant advancement in interventional medical imaging, particularly for robotic C-arm CT systems, enabling real-time, high-quality CBCT reconstructions for customized orbits. It offers a transformative solution for clinical applications requiring computational efficiency and precision in imaging.
- 20 March 2025
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 70 (7) , 075005
Open AccessDRACO: differentiable reconstruction for arbitrary CBCT orbits
Objective. This study introduces a novel method for reconstructing cone beam computed tomography (CBCT) images for arbitrary orbits, addressing the computational and memory challenges associated with traditional iterative reconstruction algorithms. Approach. The proposed method employs a differentiable shift-variant filtered backprojection neural network, optimized for arbitrary trajectories. By integrating known operators into the learning model, the approach minimizes the number of trainable parameters while enhancing model interpretability. This framework adapts seamlessly to specific orbit geometries, including non-continuous trajectories such as circular-plus-arc or sinusoidal paths, enabling faster and more accurate CBCT reconstructions. Main results. Experimental validation demonstrates that the method significantly accelerates reconstruction, reducing computation time by over 97% compared to conventional iterative algorithms. It achieves superior or comparable image quality with reduced noise, as evidenced by a 38.6% reduction in mean squared error, a 7.7% increase in peak signal-to-noise ratio, and a 5.0% improvement in the structural similarity index measure. The flexibility and robustness of the approach are confirmed through its ability to handle data from diverse scan geometries. Significance. This method represents a significant advancement in interventional medical imaging, particularly for robotic C-arm CT systems, enabling real-time, high-quality CBCT reconstructions for customized orbits. It offers a transformative solution for clinical applications requiring computational efficiency and precision in imaging.