Yipeng Hu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yipeng Hu has a total of 273 co-authors. Here are the top co-authors by number of shared publications:
- Dean C. Barratt
University College London
30 shared publications - Shaheer Ullah Saeed
University College London
22 shared publications - Matthew J. Clarkson
UCL Hawkes Institute, University College London, London, UK
19 shared publications - Mark Emberton
Division of Surgery and Interventional Science, University College London, London WC1E 6BT, UK
14 shared publications - Qianye Yang
University of Oxford
14 shared publications Dean Barratt
University College London
13 shared publications
Latest Publications
- Open Access
- 23 June 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Reasoning in machine vision by learning fast and slow thinking
Reasoning is a hallmark of human intelligence, enabling adaptive decision-making in complex unfamiliar scenarios. In contrast, machine intelligence remains bound to training data, unable to dynamically refine solutions at inference. While recent advances have explored machine reasoning - trading inference-time compute for improved performance - they focus on verbal domains such as mathematical problem-solving where explicit rules govern step-by-step solution generation. Many tasks lack sufficient labelled data and require alternative performance improvement mechanisms, such as inference-time compute. Here we present a paradigm for machine reasoning in vision, enabling performance improvements with increasing thinking time (inference-time compute), even with limited labelled data. Our approach is inspired by dual-process theories of human cognition, integrating a fast-thinking System I module for generating and verifying solutions in familiar tasks, with a slow-thinking System II module that iteratively refines predictions using self-play reinforcement learning, even when task-specific data is limited. This paradigm involves proposing, competing over, and refining solutions until convergence. We demonstrate that extended inference-time compute yields superior performance compared to large-scale supervised learning, foundation models, and human experts in vision tasks. These include computer-vision benchmarks and cancer localisation across five organs, highlighting the potential of inference-time compute for data-scarce problems.
- 23 June 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessReasoning in machine vision by learning fast and slow thinking
Reasoning is a hallmark of human intelligence, enabling adaptive decision-making in complex unfamiliar scenarios. In contrast, machine intelligence remains bound to training data, unable to dynamically refine solutions at inference. While recent advances have explored machine reasoning - trading inference-time compute for improved performance - they focus on verbal domains such as mathematical problem-solving where explicit rules govern step-by-step solution generation. Many tasks lack sufficient labelled data and require alternative performance improvement mechanisms, such as inference-time compute. Here we present a paradigm for machine reasoning in vision, enabling performance improvements with increasing thinking time (inference-time compute), even with limited labelled data. Our approach is inspired by dual-process theories of human cognition, integrating a fast-thinking System I module for generating and verifying solutions in familiar tasks, with a slow-thinking System II module that iteratively refines predictions using self-play reinforcement learning, even when task-specific data is limited. This paradigm involves proposing, competing over, and refining solutions until convergence. We demonstrate that extended inference-time compute yields superior performance compared to large-scale supervised learning, foundation models, and human experts in vision tasks. These include computer-vision benchmarks and cancer localisation across five organs, highlighting the potential of inference-time compute for data-scarce problems.
- Open Access
- 13 May 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Development and validation of a versatile foundation model for cine cardiac magnetic resonance image analysis
Background: Cardiac magnetic resonance imaging is central to cardiovascular diagnosis and management, yet extracting key clinical measurements remains time-consuming, subjective, and of limited reproducibility. Current deep learning methods often require a separate model trained from scratch for each task, and generating sufficient labelled training data demands substantial clinical expertise. Methods: We developed CineMA, a multi-view conv-transformer masked autoencoder foundation model, pre-trained on 15 million cine cardiac magnetic resonance images from 74,916 studies. The model was fine-tuned and evaluated on eight independent datasets for segmentation, landmark localisation, disease diagnosis, and prognostication, representing the largest such benchmark to date. Performance was compared against convolutional neural network baselines, including nnUNet. Results: Here we show, without dataset-specific hyperparameter tuning, CineMA approaches nnUNet performance in ventricle segmentation and ejection fraction estimation while achieving higher consistency across repeated scans. CineMA surpasses convolutional baselines in cardiovascular disease detection with notably improved specificity, and matches their performance in long-axis function measurement. Beyond cardiac diseases, CineMA shows potential for predicting systemic conditions and survival outcomes, with comparable performance across demographic subgroups. Conclusions: CineMA demonstrates accuracy, learning efficiency, adaptability, and fairness across diverse cardiac image analysis tasks, offering a strong alternative to task-specific model training for automated cardiac image analysis.
- 13 May 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessDevelopment and validation of a versatile foundation model for cine cardiac magnetic resonance image analysis
Background: Cardiac magnetic resonance imaging is central to cardiovascular diagnosis and management, yet extracting key clinical measurements remains time-consuming, subjective, and of limited reproducibility. Current deep learning methods often require a separate model trained from scratch for each task, and generating sufficient labelled training data demands substantial clinical expertise. Methods: We developed CineMA, a multi-view conv-transformer masked autoencoder foundation model, pre-trained on 15 million cine cardiac magnetic resonance images from 74,916 studies. The model was fine-tuned and evaluated on eight independent datasets for segmentation, landmark localisation, disease diagnosis, and prognostication, representing the largest such benchmark to date. Performance was compared against convolutional neural network baselines, including nnUNet. Results: Here we show, without dataset-specific hyperparameter tuning, CineMA approaches nnUNet performance in ventricle segmentation and ejection fraction estimation while achieving higher consistency across repeated scans. CineMA surpasses convolutional baselines in cardiovascular disease detection with notably improved specificity, and matches their performance in long-axis function measurement. Beyond cardiac diseases, CineMA shows potential for predicting systemic conditions and survival outcomes, with comparable performance across demographic subgroups. Conclusions: CineMA demonstrates accuracy, learning efficiency, adaptability, and fairness across diverse cardiac image analysis tasks, offering a strong alternative to task-specific model training for automated cardiac image analysis.
- Open Access
- 8 May 2026
- Published by MDPI AG in Diagnostics
- Vol. 16 (10) , 1441
Hybrid Curriculum Learning for Data-Efficient Lung Nodule Detection with YOLOv11
Background/Objectives: Accurate detection of pulmonary nodules on chest CT is critical for lung cancer screening, yet training robust detectors remains challenging due to the high cost of reliable annotations. In this work, we present a systematic study of curriculum learning for CT-based lung nodule detection on the enhanced LUNA25 benchmark and propose a hybrid curriculum learning framework for data-efficient optimization. Methods: Our approach estimates sample difficulty by fusing clinically interpretable handcrafted factors such as nodule size and count with model-driven signals such as prediction confidence from a teacher model, and constructs a three-stage progressive training curriculum from easy to hard samples. Using YOLOv11s as a strong baseline, the proposed hybrid curriculum is compared against conventional training without curriculum learning and against single-source curricula. Results: On the held-out LUNA25 test set, hybrid curriculum learning increases from 0.672 to 0.696, from 0.369 to 0.385, recall from 0.588 to 0.634, and precision from 0.725 to 0.764. Extensive data-efficiency experiments with proportional reductions (1/2, 1/5, 1/10) and fixed training samples (5000–40,000 slices) further confirm consistent gains across limited-data regimes. Conclusions: These results demonstrate that jointly leveraging intrinsic image complexity and optimization-aware feedback provides effective sample scheduling for robust and data-efficient lung nodule detection.- 8 May 2026
- Published by MDPI AG in Diagnostics
- Vol. 16 (10) , 1441
Open AccessHybrid Curriculum Learning for Data-Efficient Lung Nodule Detection with YOLOv11
Background/Objectives: Accurate detection of pulmonary nodules on chest CT is critical for lung cancer screening, yet training robust detectors remains challenging due to the high cost of reliable annotations. In this work, we present a systematic study of curriculum learning for CT-based lung nodule detection on the enhanced LUNA25 benchmark and propose a hybrid curriculum learning framework for data-efficient optimization. Methods: Our approach estimates sample difficulty by fusing clinically interpretable handcrafted factors such as nodule size and count with model-driven signals such as prediction confidence from a teacher model, and constructs a three-stage progressive training curriculum from easy to hard samples. Using YOLOv11s as a strong baseline, the proposed hybrid curriculum is compared against conventional training without curriculum learning and against single-source curricula. Results: On the held-out LUNA25 test set, hybrid curriculum learning increases from 0.672 to 0.696, from 0.369 to 0.385, recall from 0.588 to 0.634, and precision from 0.725 to 0.764. Extensive data-efficiency experiments with proportional reductions (1/2, 1/5, 1/10) and fixed training samples (5000–40,000 slices) further confirm consistent gains across limited-data regimes. Conclusions: These results demonstrate that jointly leveraging intrinsic image complexity and optimization-aware feedback provides effective sample scheduling for robust and data-efficient lung nodule detection. - 17 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 35 (10577149) , 4383-4396
PromptReg: Interactive Registration by “Corresponding Prompts” for Segment Anything Model (SAM)
Effectively establishing correspondence between two images is at the centre of image registration methods. Spatially omnipresent representations, including dense displacement fields (DDFs) and spatial (non-)rigid transformations, have been used to parameterise such correspondence. Alternatively, region-based representation uses paired regions of interest (ROIs) to represent region-level correspondence, while retaining its local and dense representation capability at pixel/voxel level if required. Thus, registration can be re-envisioned as a problem of segmenting corresponding paired ROIs in the to-be-registered images. In this work, we utilize models such as SAM, which are pre-trained on substantive datasets, to segment ROIs of the same class from two images, for a new training-free, non-iterative registration algorithm. First, a “corresponding prompt problem” is posed to find a corresponding Prompt Y on Image Y, given any vision Prompt X on Image X, such that the two respectively prompt-conditioned segmentations are a pair of corresponding ROIs from the two images. Second, we propose an “inverse prompt” solution to the corresponding prompt problem, by inverting Prompt X to the Image Y prompt space, where the Jacobian of prototypical features is used. Third, we propose a new registration algorithm that identifies multiple paired corresponding ROIs, by marginalizing the inverted Prompt X over both prompt and spatial spaces, random sampling Prompt X and spatial warping Image X. Comprehensive experiments were conducted on five applications of registering 3D prostate MR, 3D abdomen CT, 3D lung CT, 2D histopathology and, as a non-medical example, 2D aerial images. Based on metrics including Dice and target registration errors on anatomical structures, the proposed registration outperforms both intensity-based iterative algorithms and learning-based networks, even yielding competitive performance with weakly-supervised registration which requires fully-segmented training data.
- 17 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 35 (10577149) , 4383-4396
PromptReg: Interactive Registration by “Corresponding Prompts” for Segment Anything Model (SAM)
Effectively establishing correspondence between two images is at the centre of image registration methods. Spatially omnipresent representations, including dense displacement fields (DDFs) and spatial (non-)rigid transformations, have been used to parameterise such correspondence. Alternatively, region-based representation uses paired regions of interest (ROIs) to represent region-level correspondence, while retaining its local and dense representation capability at pixel/voxel level if required. Thus, registration can be re-envisioned as a problem of segmenting corresponding paired ROIs in the to-be-registered images. In this work, we utilize models such as SAM, which are pre-trained on substantive datasets, to segment ROIs of the same class from two images, for a new training-free, non-iterative registration algorithm. First, a “corresponding prompt problem” is posed to find a corresponding Prompt Y on Image Y, given any vision Prompt X on Image X, such that the two respectively prompt-conditioned segmentations are a pair of corresponding ROIs from the two images. Second, we propose an “inverse prompt” solution to the corresponding prompt problem, by inverting Prompt X to the Image Y prompt space, where the Jacobian of prototypical features is used. Third, we propose a new registration algorithm that identifies multiple paired corresponding ROIs, by marginalizing the inverted Prompt X over both prompt and spatial spaces, random sampling Prompt X and spatial warping Image X. Comprehensive experiments were conducted on five applications of registering 3D prostate MR, 3D abdomen CT, 3D lung CT, 2D histopathology and, as a non-medical example, 2D aerial images. Based on metrics including Dice and target registration errors on anatomical structures, the proposed registration outperforms both intensity-based iterative algorithms and learning-based networks, even yielding competitive performance with weakly-supervised registration which requires fully-segmented training data.
- 9 April 2026
- Published by Taylor & Francis
- p. 474-490
Medical image analysis
- 18 March 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 111, 104048
SEQUAL: Self-refining and effective querying active learning with pseudo label divergence score for carotid intima-media segmentation in ultrasound
- 18 March 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 111, 104048
SEQUAL: Self-refining and effective querying active learning with pseudo label divergence score for carotid intima-media segmentation in ultrasound
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 45 (6) , 2822-2837
Biomechanics-Informed Non-Rigid Medical Image Registration With Elasticity Theories
Biomechanical modelling of soft tissue provides a method for constraining medical image registration, such that the estimated spatial transformation is considered biophysically plausible. Existing methods either directly optimize the loss function containing the biomechanical-constrained regularization term over deformations, which takes much computational time, or are trained using biomechanically plausible data generated via finite element simulation, which is cumbersome. This work first instantiates the recently-proposed physics-informed neural networks (PINNs) to 3D elastic models that are used to establish the partial differential equations (PDEs) representing physics laws of biomechanical constraints to be satisfied. The registration algorithm that aligns point sets considering PINN-imposed biomechanics (i.e., the forward problem) is then formulated. In addition, the inverse problem and its algorithm of physical parameter (i.e., material property) estimation along with the registration are also formulated and developed. We carefully compare linear and nonlinear elasticity theories’ capabilities in solving both tasks of forward registration and inverse physical parameter identification under PINNs respectively. Furthermore, two specific network configurations that leverage one common branch or two individual branches to predict deformation vectors and biomechanical states are also constructed and compared. The proposed PINNs-based registration approaches have been extensively evaluated with three experiments, that is single and multiple patient MRI-US registration using clinical MRI-US pairs, and registration using pairs of undeformed MR images from clinical cases of prostate cancer biopsy and deformed counterparts with finite-element-computed ground-truth deformation. Results demonstrate that the proposed methods achieve state-of-the-art performances compared to biomechanical-model-based and learning-based registration approaches, and the biomechanical constraints of soft tissues have been successfully warranted after registration. The codes are available at https://github.com/ZheMin-1992/Registration_PINNs.
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 45 (6) , 2822-2837
Biomechanics-Informed Non-Rigid Medical Image Registration With Elasticity Theories
Biomechanical modelling of soft tissue provides a method for constraining medical image registration, such that the estimated spatial transformation is considered biophysically plausible. Existing methods either directly optimize the loss function containing the biomechanical-constrained regularization term over deformations, which takes much computational time, or are trained using biomechanically plausible data generated via finite element simulation, which is cumbersome. This work first instantiates the recently-proposed physics-informed neural networks (PINNs) to 3D elastic models that are used to establish the partial differential equations (PDEs) representing physics laws of biomechanical constraints to be satisfied. The registration algorithm that aligns point sets considering PINN-imposed biomechanics (i.e., the forward problem) is then formulated. In addition, the inverse problem and its algorithm of physical parameter (i.e., material property) estimation along with the registration are also formulated and developed. We carefully compare linear and nonlinear elasticity theories’ capabilities in solving both tasks of forward registration and inverse physical parameter identification under PINNs respectively. Furthermore, two specific network configurations that leverage one common branch or two individual branches to predict deformation vectors and biomechanical states are also constructed and compared. The proposed PINNs-based registration approaches have been extensively evaluated with three experiments, that is single and multiple patient MRI-US registration using clinical MRI-US pairs, and registration using pairs of undeformed MR images from clinical cases of prostate cancer biopsy and deformed counterparts with finite-element-computed ground-truth deformation. Results demonstrate that the proposed methods achieve state-of-the-art performances compared to biomechanical-model-based and learning-based registration approaches, and the biomechanical constraints of soft tissues have been successfully warranted after registration. The codes are available at https://github.com/ZheMin-1992/Registration_PINNs.
- Open Access
- 19 February 2026
- Published by Wiley in Medical Physics
- Vol. 53 (2) , e70343
On the degrees of freedom of gridded control points in learning‐based medical image registration
Background: Many registration problems are ill‐posed in homogeneous/noisy regions, and dense voxel‐wise decoders can be unnecessarily high‐dimensional. A sparse control‐point parameterization provides a compact, smooth deformation representation while reducing memory and improving stability. Purpose: This work investigates the required control points for learning‐based registration network development. In particular, as sparse as control points are configured and compared with alternative approaches, including those using scattered control points and displacements sampled at every voxel, that is, dense displacement fields. Method: We present GridReg, a learning‐based registration framework that replaces dense voxel‐wise decoding with displacement predictions at a sparse grid of control points. This design substantially cuts the parameter count and memory while retaining registration accuracy. Multiscale 3D encoder feature maps are flattened into a 1D token sequence with positional encoding to retain spatial context. The model then predicts a sparse gridded deformation field using a cross‐attention module: Each control point attends to encoder tokens within its local grid neighborhood to estimate its displacement, which is subsequently interpolated to a dense field. We further introduce grid‐adaptive training, enabling an adaptive model to operate at multiple grid sizes at inference without retraining. Results: This work quantitatively demonstrates the benefits of using sparse grids. Using three data sets for registering prostate gland, pelvic organs and neurological structures, the experimental results suggest a much improved computational efficiency, due to the prediction of sparse‐grid‐sampled displacements. Alternatively, the superior registration performance was obtained using the proposed approach, with the similiar or less compute cost, compared with existing algorithms that predict DDFs (e.g., VoxelMorph/TransMorph) or displacements sampled on scattered key points (KeyMorph). Conclusion: We conclude that predicting sparsely gridded displacements provides reduced computational cost and/or improved performance, independent of the encoder architecture, and can be readily implemented. Therefore, GridReg should potentially be considered for many registration tasks with adaptive grid sizes. The code is available via git@github.com:yanwenCi/GridReg.git .
- 19 February 2026
- Published by Wiley in Medical Physics
- Vol. 53 (2) , e70343
Open AccessOn the degrees of freedom of gridded control points in learning‐based medical image registration
Background: Many registration problems are ill‐posed in homogeneous/noisy regions, and dense voxel‐wise decoders can be unnecessarily high‐dimensional. A sparse control‐point parameterization provides a compact, smooth deformation representation while reducing memory and improving stability. Purpose: This work investigates the required control points for learning‐based registration network development. In particular, as sparse as control points are configured and compared with alternative approaches, including those using scattered control points and displacements sampled at every voxel, that is, dense displacement fields. Method: We present GridReg, a learning‐based registration framework that replaces dense voxel‐wise decoding with displacement predictions at a sparse grid of control points. This design substantially cuts the parameter count and memory while retaining registration accuracy. Multiscale 3D encoder feature maps are flattened into a 1D token sequence with positional encoding to retain spatial context. The model then predicts a sparse gridded deformation field using a cross‐attention module: Each control point attends to encoder tokens within its local grid neighborhood to estimate its displacement, which is subsequently interpolated to a dense field. We further introduce grid‐adaptive training, enabling an adaptive model to operate at multiple grid sizes at inference without retraining. Results: This work quantitatively demonstrates the benefits of using sparse grids. Using three data sets for registering prostate gland, pelvic organs and neurological structures, the experimental results suggest a much improved computational efficiency, due to the prediction of sparse‐grid‐sampled displacements. Alternatively, the superior registration performance was obtained using the proposed approach, with the similiar or less compute cost, compared with existing algorithms that predict DDFs (e.g., VoxelMorph/TransMorph) or displacements sampled on scattered key points (KeyMorph). Conclusion: We conclude that predicting sparsely gridded displacements provides reduced computational cost and/or improved performance, independent of the encoder architecture, and can be readily implemented. Therefore, GridReg should potentially be considered for many registration tasks with adaptive grid sizes. The code is available via git@github.com:yanwenCi/GridReg.git .
- 14 February 2026
- Published by Springer Nature
- p. 64-74
Radiomic Fingerprints for Knee MR Image Assessment
- 14 February 2026
- Published by Springer Nature
- p. 64-74
Radiomic Fingerprints for Knee MR Image Assessment
- Open Access
- 4 February 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
A computationally frugal, open-source chest CT foundation model for thoracic disease detection in lung cancer screening programmes
Background: Low-dose computed tomography (LDCT) employed in lung cancer screening (LCS) programmes is increasing in uptake worldwide. LCS programmes herald a generational opportunity to simultaneously detect cancer and non-cancer-related early-stage lung disease, yet these efforts are hampered by a shortage of radiologists to interpret scans at scale. Here, we present TANGERINE, a computationally frugal, open-source vision foundation model for volumetric LDCT analysis. Methods: Designed for broad accessibility and rapid adaptation, TANGERINE can be fine-tuned off the shelf for a wide range of disease-specific tasks with limited computational resources and training data. The model is pretrained using self-supervised learning on more than 98,000 thoracic LDCT scans, including the United Kingdom’s largest LCS initiative to date and 27 public datasets. By extending a masked autoencoder framework to three-dimensional imaging, TANGERINE provides a scalable solution for LDCT analysis, combining architectural simplicity, public availability, and modest computational requirements. Results: TANGERINE demonstrates superior computational and data efficiency in a retrospective multi-dataset analysis: it converges rapidly during fine-tuning, requiring significantly fewer graphics processing unit hours than models trained from scratch, and achieves comparable or superior performance using only a fraction of the fine-tuning data. The model achieves strong performance across 14 disease classification tasks, including lung cancer and multiple respiratory diseases, and generalises robustly across diverse clinical centres. Conclusions: TANGERINE’s accessible, open-source, lightweight design lays the foundation for rapid integration into next-generation medical imaging tools, enabling lung cancer screening programmes to pivot from a singular focus on lung cancer detection toward comprehensive respiratory disease management in high-risk populations.
- 4 February 2026
- Published by Springer Nature in Communications Medicine
- Vol. 6 (1)
Open AccessA computationally frugal, open-source chest CT foundation model for thoracic disease detection in lung cancer screening programmes
Background: Low-dose computed tomography (LDCT) employed in lung cancer screening (LCS) programmes is increasing in uptake worldwide. LCS programmes herald a generational opportunity to simultaneously detect cancer and non-cancer-related early-stage lung disease, yet these efforts are hampered by a shortage of radiologists to interpret scans at scale. Here, we present TANGERINE, a computationally frugal, open-source vision foundation model for volumetric LDCT analysis. Methods: Designed for broad accessibility and rapid adaptation, TANGERINE can be fine-tuned off the shelf for a wide range of disease-specific tasks with limited computational resources and training data. The model is pretrained using self-supervised learning on more than 98,000 thoracic LDCT scans, including the United Kingdom’s largest LCS initiative to date and 27 public datasets. By extending a masked autoencoder framework to three-dimensional imaging, TANGERINE provides a scalable solution for LDCT analysis, combining architectural simplicity, public availability, and modest computational requirements. Results: TANGERINE demonstrates superior computational and data efficiency in a retrospective multi-dataset analysis: it converges rapidly during fine-tuning, requiring significantly fewer graphics processing unit hours than models trained from scratch, and achieves comparable or superior performance using only a fraction of the fine-tuning data. The model achieves strong performance across 14 disease classification tasks, including lung cancer and multiple respiratory diseases, and generalises robustly across diverse clinical centres. Conclusions: TANGERINE’s accessible, open-source, lightweight design lays the foundation for rapid integration into next-generation medical imaging tools, enabling lung cancer screening programmes to pivot from a singular focus on lung cancer detection toward comprehensive respiratory disease management in high-risk populations.