Zeju Li

Chen Chen, Zeju Li, Cheng Ouyang et al.

Convolutional neural networks (CNNs) have achieved remarkable segmentation accuracy on benchmark datasets where training and test sets are from the same domain, yet their performance can degrade significantly on unseen domains, which hinders the deployment of CNNs in many clinical scenarios. Most existing works improve model out-of-domain (OOD) robustness by collecting multi-domain datasets for training, which is expensive and may not always be feasible due to privacy and logistical issues. In this work, we focus on improving model robustness using a single-domain dataset only. We propose a novel data augmentation framework called MaxStyle, which maximizes the effectiveness of style augmentation for model OOD performance. It attaches an auxiliary style-augmented image decoder to a segmentation network for robust feature learning and data augmentation. Importantly, MaxStyle augments data with improved image style diversity and hardness, by expanding the style space with noise and searching for the worst-case style composition of latent features via adversarial training. With extensive experiments on multiple public cardiac and prostate MR datasets, we demonstrate that MaxStyle leads to significantly improved out-of-distribution robustness against unseen corruptions as well as common distribution shifts across multiple, different, unseen sites and unknown image sequences under both low- and high-training data settings. The code can be found at https://github.com/cherise215/MaxStyle.

Zeju Li, Konstantinos Kamnitsas, Cheng Ouyang et al.

Background samples provide key contextual information for segmenting regions of interest (ROIs). However, they always cover a diverse set of structures, causing difficulties for the segmentation model to learn good decision boundaries with high sensitivity and precision. The issue concerns the highly heterogeneous nature of the background class, resulting in multi-modal distributions. Empirically, we find that neural networks trained with heterogeneous background struggle to map the corresponding contextual samples to compact clusters in feature space. As a result, the distribution over background logit activations may shift across the decision boundary, leading to systematic over-segmentation across different datasets and tasks. In this study, we propose context label learning (CoLab) to improve the context representations by decomposing the background class into several subclasses. Specifically, we train an auxiliary network as a task generator, along with the primary segmentation model, to automatically generate context labels that positively affect the ROI segmentation accuracy. Extensive experiments are conducted on several challenging segmentation tasks and datasets. The results demonstrate that CoLab can guide the segmentation model to map the logits of background samples away from the decision boundary, resulting in significantly improved segmentation accuracy. Code is available.

Felix Wagner, Zeju Li, Pramit Saha et al.

Deployment of Deep Neural Networks in medical imaging is hindered by distribution shift between training data and data processed after deployment, causing performance degradation. Post-Deployment Adaptation (PDA) addresses this by tailoring a pre-trained, deployed model to the target data distribution using limited labelled or entirely unlabelled target data, while assuming no access to source training data as they cannot be deployed with the model due to privacy concerns and their large size. This makes reliable adaptation challenging due to limited learning signal. This paper challenges this assumption and introduces FedPDA, a novel adaptation framework that brings the utility of learning from remote data from Federated Learning into PDA. FedPDA enables a deployed model to obtain information from source data via remote gradient exchange, while aiming to optimize the model specifically for the target domain. Tailored for FedPDA, we introduce a novel optimization method StarAlign (Source-Target Remote Gradient Alignment) that aligns gradients between source-target domain pairs by maximizing their inner product, to facilitate learning a target-specific model. We demonstrate the method's effectiveness using multi-center databases for the tasks of cancer metastases detection and skin lesion classification, where our method compares favourably to previous work. Code is available at: https://github.com/FelixWag/StarAlign

Xiao Gu, Yao Guo, Zeju Li et al. · oxford

Machine learning models fail to perform well on real-world applications when 1) the category distribution P(Y) of the training dataset suffers from long-tailed distribution and 2) the test data is drawn from different conditional distributions P(X|Y). Existing approaches cannot handle the scenario where both issues exist, which however is common for real-world applications. In this study, we took a step forward and looked into the problem of long-tailed classification under domain shifts. We designed three novel core functional blocks including Distribution Calibrated Classification Loss, Visual-Semantic Mapping and Semantic-Similarity Guided Augmentation. Furthermore, we adopted a meta-learning framework which integrates these three blocks to improve domain generalization on unseen target domains. Two new datasets were proposed for this problem, named AWA2-LTS and ImageNet-LTS. We evaluated our method on the two datasets and extensive experimental results demonstrate that our proposed method can achieve superior performance over state-of-the-art long-tailed/domain generalization approaches and the combinations. Source codes and datasets can be found at our project page https://xiaogu.site/LTDS.

Xiangyu Wen, Yuang Zhao, Xiaoyu Xu et al.

The transition of agentic AI from brittle prototypes to production systems is stalled by a pervasive crisis of craft. We suggest that the prevailing orchestration paradigm-delegating the system control loop to large language models and merely patching with heuristic guardrails-is the root cause of this fragility. Instead, we propose Arbiter-K, a Governance-First execution architecture that reconceptualizes the underlying model as a Probabilistic Processing Unit encapsulated by a deterministic, neuro-symbolic kernel. Arbiter-K implements a Semantic Instruction Set Architecture (ISA) to reify probabilistic messages into discrete instructions. This allows the kernel to maintain a Security Context Registry and construct an Instruction Dependency Graph at runtime, enabling active taint propagation based on the data-flow pedigree of each reasoning node. By leveraging this mechanism, Arbiter-K precisely interdicts unsafe trajectories at deterministic sinks (e.g., high-risk tool calls or unauthorized network egress) and enables autonomous execution correction and architectural rollback when security policies are triggered. Evaluations on OpenClaw and NanoBot demonstrate that Arbiter-K enforces security as a microarchitectural property, achieving 76% to 95% unsafe interception for a 92.79% absolute gain over native policies. The code is publicly available at https://github.com/cure-lab/ArbiterOS.

Cheng Ouyang, Shuo Wang, Chen Chen et al.

Probability calibration for deep models is highly desirable in safety-critical applications such as medical imaging. It makes output probabilities of deep networks interpretable, by aligning prediction probability with the actual accuracy in test data. In image segmentation, well-calibrated probabilities allow radiologists to identify regions where model-predicted segmentations are unreliable. These unreliable predictions often occur to out-of-domain (OOD) images that are caused by imaging artifacts or unseen imaging protocols. Unfortunately, most previous calibration methods for image segmentation perform sub-optimally on OOD images. To reduce the calibration error when confronted with OOD images, we propose a novel post-hoc calibration model. Our model leverages the pixel susceptibility against perturbations at the local level, and the shape prior information at the global level. The model is tested on cardiac MRI segmentation datasets that contain unseen imaging artifacts and images from an unseen imaging protocol. We demonstrate reduced calibration errors compared with the state-of-the-art calibration algorithm.

Zeju Li, Konstantinos Kamnitsas, Mobarakol Islam et al.

Machine learning models are typically deployed in a test setting that differs from the training setting, potentially leading to decreased model performance because of domain shift. If we could estimate the performance that a pre-trained model would achieve on data from a specific deployment setting, for example a certain clinic, we could judge whether the model could safely be deployed or if its performance degrades unacceptably on the specific data. Existing approaches estimate this based on the confidence of predictions made on unlabeled test data from the deployment's domain. We find existing methods struggle with data that present class imbalance, because the methods used to calibrate confidence do not account for bias induced by class imbalance, consequently failing to estimate class-wise accuracy. Here, we introduce class-wise calibration within the framework of performance estimation for imbalanced datasets. Specifically, we derive class-specific modifications of state-of-the-art confidence-based model evaluation methods including temperature scaling (TS), difference of confidences (DoC), and average thresholded confidence (ATC). We also extend the methods to estimate Dice similarity coefficient (DSC) in image segmentation. We conduct experiments on four tasks and find the proposed modifications consistently improve the estimation accuracy for imbalanced datasets. Our methods improve accuracy estimation by 18\% in classification under natural domain shifts, and double the estimation accuracy on segmentation tasks, when compared with prior methods.

Mobarakol Islam, Zeju Li, Ben Glocker

Deep neural networks have shown impressive performance for image-based disease detection. Performance is commonly evaluated through clinical validation on independent test sets to demonstrate clinically acceptable accuracy. Reporting good performance metrics on test sets, however, is not always a sufficient indication of the generalizability and robustness of an algorithm. In particular, when the test data is drawn from the same distribution as the training data, the iid test set performance can be an unreliable estimate of the accuracy on new data. In this paper, we employ stress testing to assess model robustness and subgroup performance disparities in disease detection models. We design progressive stress testing using five different bidirectional and unidirectional image perturbations with six different severity levels. As a use case, we apply stress tests to measure the robustness of disease detection models for chest X-ray and skin lesion images, and demonstrate the importance of studying class and domain-specific model behaviour. Our experiments indicate that some models may yield more robust and equitable performance than others. We also find that pretraining characteristics play an important role in downstream robustness. We conclude that progressive stress testing is a viable and important tool and should become standard practice in the clinical validation of image-based disease detection models.

Qin Wang, Zhiqing He, Yu Liu et al.

Reliable interpretation of echocardiography (Echo) is crucial for assessing cardiac function, which demands clinicians to synchronously orchestrate multiple capabilities, including visual observation (eyes), manual measurement (hands), and expert knowledge learning and reasoning (minds). While current task-specific deep-learning approaches and multimodal large language models have demonstrated promise in assisting Echo analysis through automated segmentation or reasoning, they remain focused on restricted skills, i.e., eyes-hands or eyes-minds, thereby limiting clinical reliability and utility. To address these issues, we propose EchoAgent, an agentic system tailored for end-to-end Echo interpretation, which achieves a fully coordinated eyes-hands-minds workflow that learns, observes, operates, and reasons like a cardiac sonographer. First, we introduce an expertise-driven cognition engine where our agent can automatically assimilate credible Echo guidelines into a structured knowledge base, thus constructing an Echo-customized mind. Second, we devise a hierarchical collaboration toolkit to endow EchoAgent with eyes-hands, which can automatically parse Echo video streams, identify cardiac views, perform anatomical segmentation, and quantitative measurement. Third, we integrate the perceived multimodal evidence with the exclusive knowledge base into an orchestrated reasoning hub to conduct explainable inferences. We evaluate EchoAgent on CAMUS and MIMIC-EchoQA datasets, which cover 48 distinct echocardiographic views spanning 14 cardiac anatomical regions. Experimental results show that EchoAgent achieves optimal performance across diverse structure analyses, yielding overall accuracy of up to 80.00%. Importantly, EchoAgent empowers a single system with abilities to learn, observe, operate and reason like an echocardiologist, which holds great promise for reliable Echo interpretation.

Zeju Li, Chao Zhang, Xiaoyan Wang et al.

The remarkable potential of multi-modal large language models (MLLMs) in comprehending both vision and language information has been widely acknowledged. However, the scarcity of 3D scenes-language pairs in comparison to their 2D counterparts, coupled with the inadequacy of existing approaches in understanding of 3D scenes by LLMs, poses a significant challenge. In response, we collect and construct an extensive dataset comprising 75K instruction-response pairs tailored for 3D scenes. This dataset addresses tasks related to 3D VQA, 3D grounding, and 3D conversation. To further enhance the integration of 3D spatial information into LLMs, we introduce a novel and efficient prompt tuning paradigm, 3DMIT. This paradigm eliminates the alignment stage between 3D scenes and language and extends the instruction prompt with the 3D modality information including the entire scene and segmented objects. We evaluate the effectiveness of our method across diverse tasks in the 3D scene domain and find that our approach serves as a strategic means to enrich LLMs' comprehension of the 3D world. Our code is available at https://github.com/staymylove/3DMIT.

Zeju Li, Ying-Qiu Zheng, Chen Chen et al.

Semi-supervised learning (SSL) algorithms struggle to perform well when exposed to imbalanced training data. In this scenario, the generated pseudo-labels can exhibit a bias towards the majority class, and models that employ these pseudo-labels can further amplify this bias. Here we investigate pseudo-labeling strategies for imbalanced SSL including pseudo-label refinement and threshold adjustment, through the lens of statistical analysis. We find that existing SSL algorithms which generate pseudo-labels using heuristic strategies or uncalibrated model confidence are unreliable when imbalanced class distributions bias pseudo-labels. To address this, we introduce SEmi-supervised learning with pseudo-label optimization based on VALidation data (SEVAL) to enhance the quality of pseudo-labelling for imbalanced SSL. We propose to learn refinement and thresholding parameters from a partition of the training dataset in a class-balanced way. SEVAL adapts to specific tasks with improved pseudo-labels accuracy and ensures pseudo-labels correctness on a per-class basis. Our experiments show that SEVAL surpasses state-of-the-art SSL methods, delivering more accurate and effective pseudo-labels in various imbalanced SSL situations. SEVAL, with its simplicity and flexibility, can enhance various SSL techniques effectively. The code is publicly available (https://github.com/ZerojumpLine/SEVAL).

Yi Liu, Changran Xu, Yunhao Zhou et al.

Recent advancements in large language models (LLMs) have shown significant potential for automating hardware description language (HDL) code generation from high-level natural language instructions. While fine-tuning has improved LLMs' performance in hardware design tasks, prior efforts have largely focused on Verilog generation, overlooking the equally critical task of Verilog understanding. Furthermore, existing models suffer from weak alignment between natural language descriptions and Verilog code, hindering the generation of high-quality, synthesizable designs. To address these issues, we present DeepRTL, a unified representation model that excels in both Verilog understanding and generation. Based on CodeT5+, DeepRTL is fine-tuned on a comprehensive dataset that aligns Verilog code with rich, multi-level natural language descriptions. We also introduce the first benchmark for Verilog understanding and take the initiative to apply embedding similarity and GPT Score to evaluate the models' understanding capabilities. These metrics capture semantic similarity more accurately than traditional methods like BLEU and ROUGE, which are limited to surface-level n-gram overlaps. By adapting curriculum learning to train DeepRTL, we enable it to significantly outperform GPT-4 in Verilog understanding tasks, while achieving performance on par with OpenAI's o1-preview model in Verilog generation tasks.

Jianyuan Zhong, Zeju Li, Zhijian Xu et al.

We present Dyve, a dynamic process verifier that enhances reasoning error detection in large language models by integrating fast and slow thinking, inspired by Kahneman's Systems Theory. Dyve adaptively applies immediate token-level confirmation System 1 for straightforward steps and comprehensive analysis System 2 for complex ones. Leveraging a novel step-wise consensus-filtered process supervision technique, combining Monte Carlo estimation with LLM based evaluation, Dyve curates high-quality supervision signals from noisy data. Experimental results on ProcessBench and the MATH dataset confirm that Dyve significantly outperforms existing process-based verifiers and boosts performance in Best-of-N settings.

Mingyu Liu, Jiuhe Shu, Hui Chen et al.

A fundamental challenge in embodied intelligence is developing expressive and compact state representations for efficient world modeling and decision making. However, existing methods often fail to achieve this balance, yielding representations that are either overly redundant or lacking in task-critical information. We propose an unsupervised approach that learns a highly compressed two-token state representation using a lightweight encoder and a pre-trained Diffusion Transformer (DiT) decoder, capitalizing on its strong generative prior. Our representation is efficient, interpretable, and integrates seamlessly into existing VLA-based models, improving performance by 14.3% on LIBERO and 30% in real-world task success with minimal inference overhead. More importantly, we find that the difference between these tokens, obtained via latent interpolation, naturally serves as a highly effective latent action, which can be further decoded into executable robot actions. This emergent capability reveals that our representation captures structured dynamics without explicit supervision. We name our method StaMo for its ability to learn generalizable robotic Motion from compact State representation, which is encoded from static images, challenging the prevalent dependence to learning latent action on complex architectures and video data. The resulting latent actions also enhance policy co-training, outperforming prior methods by 10.4% with improved interpretability. Moreover, our approach scales effectively across diverse data sources, including real-world robot data, simulation, and human egocentric video.

Jianyuan Zhong, Zeju Li, Zhijian Xu et al.

Large Language Model (LLM) reasoning for complex tasks inherently involves a trade-off between solution accuracy and computational efficiency. The subsequent step of verification, while intended to improve performance, further complicates this landscape by introducing its own challenging trade-off: sophisticated Generative Reward Models (GenRMs) can be computationally prohibitive if naively integrated with LLMs at test-time, while simpler, faster methods may lack reliability. To overcome these challenges, we introduce FlexiVe, a novel generative verifier that flexibly balances computational resources between rapid, reliable fast thinking and meticulous slow thinking using a Flexible Allocation of Verification Budget strategy. We further propose the Solve-Detect-Verify pipeline, an efficient inference-time scaling framework that intelligently integrates FlexiVe, proactively identifying solution completion points to trigger targeted verification and provide focused solver feedback. Experiments show FlexiVe achieves superior accuracy in pinpointing errors within reasoning traces on ProcessBench. Furthermore, on challenging mathematical reasoning benchmarks (AIME 2024, AIME 2025, and CNMO), our full approach outperforms baselines like self-consistency in reasoning accuracy and inference efficiency. Our system offers a scalable and effective solution to enhance LLM reasoning at test time.

Peng Huang, Junhu Fu, Bowen Guo et al.

As the appearance of medical images is influenced by multiple underlying factors, generative models require rich attribute information beyond labels to produce realistic and diverse images. For instance, generating an image of skin lesion with specific patterns demands descriptions that go beyond diagnosis, such as shape, size, texture, and color. However, such detailed descriptions are not always accessible. To address this, we explore a framework, termed Visual Attribute Prompts (VAP)-Diffusion, to leverage external knowledge from pre-trained Multi-modal Large Language Models (MLLMs) to improve the quality and diversity of medical image generation. First, to derive descriptions from MLLMs without hallucination, we design a series of prompts following Chain-of-Thoughts for common medical imaging tasks, including dermatologic, colorectal, and chest X-ray images. Generated descriptions are utilized during training and stored across different categories. During testing, descriptions are randomly retrieved from the corresponding category for inference. Moreover, to make the generator robust to unseen combination of descriptions at the test time, we propose a Prototype Condition Mechanism that restricts test embeddings to be similar to those from training. Experiments on three common types of medical imaging across four datasets verify the effectiveness of VAP-Diffusion.

Zeju Li, Konstantinos Kamnitsas, Qi Dou et al.

This paper presents an effective and general data augmentation framework for medical image segmentation. We adopt a computationally efficient and data-efficient gradient-based meta-learning scheme to explicitly align the distribution of training and validation data which is used as a proxy for unseen test data. We improve the current data augmentation strategies with two core designs. First, we learn class-specific training-time data augmentation (TRA) effectively increasing the heterogeneity within the training subsets and tackling the class imbalance common in segmentation. Second, we jointly optimize TRA and test-time data augmentation (TEA), which are closely connected as both aim to align the training and test data distribution but were so far considered separately in previous works. We demonstrate the effectiveness of our method on four medical image segmentation tasks across different scenarios with two state-of-the-art segmentation models, DeepMedic and nnU-Net. Extensive experimentation shows that the proposed data augmentation framework can significantly and consistently improve the segmentation performance when compared to existing solutions. Code is publicly available.

Alessa Hering, Lasse Hansen, Tony C. W. Mok et al.

Image registration is a fundamental medical image analysis task, and a wide variety of approaches have been proposed. However, only a few studies have comprehensively compared medical image registration approaches on a wide range of clinically relevant tasks. This limits the development of registration methods, the adoption of research advances into practice, and a fair benchmark across competing approaches. The Learn2Reg challenge addresses these limitations by providing a multi-task medical image registration data set for comprehensive characterisation of deformable registration algorithms. A continuous evaluation will be possible at https://learn2reg.grand-challenge.org. Learn2Reg covers a wide range of anatomies (brain, abdomen, and thorax), modalities (ultrasound, CT, MR), availability of annotations, as well as intra- and inter-patient registration evaluation. We established an easily accessible framework for training and validation of 3D registration methods, which enabled the compilation of results of over 65 individual method submissions from more than 20 unique teams. We used a complementary set of metrics, including robustness, accuracy, plausibility, and runtime, enabling unique insight into the current state-of-the-art of medical image registration. This paper describes datasets, tasks, evaluation methods and results of the challenge, as well as results of further analysis of transferability to new datasets, the importance of label supervision, and resulting bias. While no single approach worked best across all tasks, many methodological aspects could be identified that push the performance of medical image registration to new state-of-the-art performance. Furthermore, we demystified the common belief that conventional registration methods have to be much slower than deep-learning-based methods.

Cheng Ouyang, Chen Chen, Surui Li et al.

Deep learning models usually suffer from domain shift issues, where models trained on one source domain do not generalize well to other unseen domains. In this work, we investigate the single-source domain generalization problem: training a deep network that is robust to unseen domains, under the condition that training data is only available from one source domain, which is common in medical imaging applications. We tackle this problem in the context of cross-domain medical image segmentation. Under this scenario, domain shifts are mainly caused by different acquisition processes. We propose a simple causality-inspired data augmentation approach to expose a segmentation model to synthesized domain-shifted training examples. Specifically, 1) to make the deep model robust to discrepancies in image intensities and textures, we employ a family of randomly-weighted shallow networks. They augment training images using diverse appearance transformations. 2) Further we show that spurious correlations among objects in an image are detrimental to domain robustness. These correlations might be taken by the network as domain-specific clues for making predictions, and they may break on unseen domains. We remove these spurious correlations via causal intervention. This is achieved by resampling the appearances of potentially correlated objects independently. The proposed approach is validated on three cross-domain segmentation tasks: cross-modality (CT-MRI) abdominal image segmentation, cross-sequence (bSSFP-LGE) cardiac MRI segmentation, and cross-center prostate MRI segmentation. The proposed approach yields consistent performance gains compared with competitive methods when tested on unseen domains.

Chen Chen, Chen Qin, Cheng Ouyang et al.

The success of neural networks on medical image segmentation tasks typically relies on large labeled datasets for model training. However, acquiring and manually labeling a large medical image set is resource-intensive, expensive, and sometimes impractical due to data sharing and privacy issues. To address this challenge, we propose AdvChain, a generic adversarial data augmentation framework, aiming at improving both the diversity and effectiveness of training data for medical image segmentation tasks. AdvChain augments data with dynamic data augmentation, generating randomly chained photo-metric and geometric transformations to resemble realistic yet challenging imaging variations to expand training data. By jointly optimizing the data augmentation model and a segmentation network during training, challenging examples are generated to enhance network generalizability for the downstream task. The proposed adversarial data augmentation does not rely on generative networks and can be used as a plug-in module in general segmentation networks. It is computationally efficient and applicable for both low-shot supervised and semi-supervised learning. We analyze and evaluate the method on two MR image segmentation tasks: cardiac segmentation and prostate segmentation with limited labeled data. Results show that the proposed approach can alleviate the need for labeled data while improving model generalization ability, indicating its practical value in medical imaging applications.

Zeju Li, Han Li, Hu Han et al.

Although chest X-ray (CXR) offers a 2D projection with overlapped anatomies, it is widely used for clinical diagnosis. There is clinical evidence supporting that decomposing an X-ray image into different components (e.g., bone, lung and soft tissue) improves diagnostic value. We hereby propose a decomposition generative adversarial network (DecGAN) to anatomically decompose a CXR image but with unpaired data. We leverage the anatomy knowledge embedded in CT, which features a 3D volume with clearly visible anatomies. Our key idea is to embed CT priori decomposition knowledge into the latent space of unpaired CXR autoencoder. Specifically, we train DecGAN with a decomposition loss, adversarial losses, cycle-consistency losses and a mask loss to guarantee that the decomposed results of the latent space preserve realistic body structures. Extensive experiments demonstrate that DecGAN provides superior unsupervised CXR bone suppression results and the feasibility of modulating CXR components by latent space disentanglement. Furthermore, we illustrate the diagnostic value of DecGAN and demonstrate that it outperforms the state-of-the-art approaches in terms of predicting 11 out of 14 common lung diseases.

Zeju Li, Konstantinos Kamnitsas, Ben Glocker

Overfitting in deep learning has been the focus of a number of recent works, yet its exact impact on the behavior of neural networks is not well understood. This study analyzes overfitting by examining how the distribution of logits alters in relation to how much the model overfits. Specifically, we find that when training with few data samples, the distribution of logit activations when processing unseen test samples of an under-represented class tends to shift towards and even across the decision boundary, while the over-represented class seems unaffected. In image segmentation, foreground samples are often heavily under-represented. We observe that sensitivity of the model drops as a result of overfitting, while precision remains mostly stable. Based on our analysis, we derive asymmetric modifications of existing loss functions and regularizers including a large margin loss, focal loss, adversarial training and mixup, which specifically aim at reducing the shift observed when embedding unseen samples of the under-represented class. We study the case of binary segmentation of brain tumor core and show that our proposed simple modifications lead to significantly improved segmentation performance over the symmetric variants.

Wenjun Yan, Yuanyuan Wang, Zeju Li et al.

Quantitative assessment of left ventricle (LV) function from cine MRI has significant diagnostic and prognostic value for cardiovascular disease patients. The temporal movement of LV provides essential information on the contracting/relaxing pattern of heart, which is keenly evaluated by clinical experts in clinical practice. Inspired by the expert way of viewing Cine MRI, we propose a new CNN module that is able to incorporate the temporal information into LV segmentation from cine MRI. In the proposed CNN, the optical flow (OF) between neighboring frames is integrated and aggregated at feature level, such that temporal coherence in cardiac motion can be taken into account during segmentation. The proposed module is integrated into the U-net architecture without need of additional training. Furthermore, dilated convolution is introduced to improve the spatial accuracy of segmentation. Trained and tested on the Cardiac Atlas database, the proposed network resulted in a Dice index of 95% and an average perpendicular distance of 0.9 pixels for the middle LV contour, significantly outperforming the original U-net that processes each frame individually. Notably, the proposed method improved the temporal coherence of LV segmentation results, especially at the LV apex and base where the cardiac motion is difficult to follow.