Zhihui Chen

Zhihui Chen, Mengling Feng

Medical image editing has emerged as a pivotal technology with broad applications in data augmentation, model interpretability, medical education, and treatment simulation. However, the lack of large-scale, high-quality, and openly accessible datasets tailored for medical contexts with strict anatomical and clinical constraints has significantly hindered progress in this domain. To bridge this gap, we introduce Med-Banana-50K, a comprehensive dataset of over 50k medically curated image edits spanning chest X-ray, brain MRI, and fundus photography across 23 diseases. Each sample supports bidirectional lesion editing (addition and removal) and is constructed using Gemini-2.5-Flash-Image based on real clinical images. A key differentiator of our dataset is the medically grounded quality control protocol: we employ an LLM-as-Judge evaluation framework with criteria such as instruction compliance, structural plausibility, image realism, and fidelity preservation, alongside iterative refinement over up to five rounds. Additionally, Med-Banana-50K includes around 37,000 failed editing attempts with full evaluation logs to support preference learning and alignment research. By offering a large-scale, medically rigorous, and fully documented resource, Med-Banana-50K establishes a critical foundation for developing and evaluating reliable medical image editing systems. Our dataset and code are publicly available. [https://github.com/richardChenzhihui/med-banana-50k].

Qianyi Xu, Gousia Habib, Feng Wu et al.

Reinforcement Learning (RL) offers a powerful framework for optimizing dynamic treatment regimes (DTRs). However, clinical RL is fundamentally bottlenecked by reward engineering: the challenge of defining signals that safely and effectively guide policy learning in complex, sparse offline environments. Existing approaches often rely on manual heuristics that fail to generalize across diverse pathologies. To address this, we propose an automated pipeline leveraging Large Language Models (LLMs) for offline reward design and verification. We formulate the reward function using potential functions consisted of three core components: survival, confidence, and competence. We further introduce quantitative metrics to rigorously evaluate and select the optimal reward structure prior to deployment. By integrating LLM-driven domain knowledge, our framework automates the design of reward functions for specific diseases while significantly enhancing the performance of the resulting policies.

Qi Cai, Jingwen Chen, Chengmin Gao et al.

The evolution of visual generative models has long been constrained by fragmented architectures relying on disjoint text encoders and external VAEs. In this report, we present HiDream-O1-Image, a natively unified generative foundation model via pixel-space Diffusion Transformer, that pioneers a paradigm shift from modular architectures to an end-to-end in-context visual generation engine. By mapping raw image pixels, text tokens, and task-specific conditions into a single shared token space, HiDream-O1-Image achieves a structural unification of multimodal inputs within an Unified Transformer (UiT) architecture. This native encoding paradigm eliminates the need for separate VAEs or disjoint pre-trained text encoders, allowing the model to treat diverse generation and editing tasks as a consistent in-context reasoning process. Extensive experiments show that HiDream-O1-Image excels across various generation tasks, including text-to-image generation, instruction-based editing, and subject-driven personalization. Notably, with only 8B parameters, HiDream-O1-Image (8B) achieves performance parity with or even surpasses established state-of-the-art models with significantly larger parameters (e.g., 27B Qwen-Image). Crucially, to validate the immense scalability of this paradigm, we successfully scale the architecture up to over 200B parameters. Experimental results demonstrate that this massive-scale version HiDream-O1-Image-Pro (200B+) unlocks unprecedented generative capabilities and superior performance, establishing new state-of-the-art benchmarks. Ultimately, HiDream-O1-Image highlights the immense potential of natively unified architectures and charts a highly scalable path toward next-generation multimodal AI.

Zhihui Chen, Kai He, Qingyuan Lei et al.

Text-guided image editors can now manipulate authentic medical scans with high fidelity, enabling lesion implantation/removal that threatens clinical trust and safety. Existing defenses are inadequate for healthcare. Medical detectors are largely black-box, while MLLM-based explainers are typically post-hoc, lack medical expertise, and may hallucinate evidence on ambiguous cases. We present MedForge, a data-and-method solution for pre-hoc, evidence-grounded medical forgery detection. We introduce MedForge-90K, a large-scale benchmark of realistic lesion edits across 19 pathologies with expert-guided reasoning supervision via doctor inspection guidelines and gold edit locations. Building on it, MedForge-Reasoner performs localize-then-analyze reasoning, predicting suspicious regions before producing a verdict, and is further aligned with Forgery-aware GSPO to strengthen grounding and reduce hallucinations. Experiments demonstrate state-of-the-art detection accuracy and trustworthy, expert-aligned explanations.

Qiyuan Wang, Zhihui Chen, Chenye Wu

While the advanced machine learning algorithms are effective in load forecasting, they often suffer from low data utilization, and hence their superior performance relies on massive datasets. This motivates us to design machine learning algorithms with improved data utilization. Specifically, we consider the load forecasting for a new user in the system by observing only few shots (data points) of its energy consumption. This task is challenging since the limited samples are insufficient to exploit the temporal characteristics, essential for load forecasting. Nonetheless, we notice that there are not too many temporal characteristics for residential loads due to the limited kinds of human lifestyle. Hence, we propose to utilize the historical load profile data from existing users to conduct effective clustering, which mitigates the challenges brought by the limited samples. Specifically, we first design a feature extraction clustering method for categorizing historical data. Then, inheriting the prior knowledge from the clustering results, we propose a two-phase Long Short Term Memory (LSTM) model to conduct load forecasting for new users. The proposed method outperforms the traditional LSTM model, especially when the training sample size fails to cover a whole period (i.e., 24 hours in our task). Extensive case studies on two real-world datasets and one synthetic dataset verify the effectiveness and efficiency of our method.

Qiangqiang Wu, Zhihui Chen, Lin Cheng et al.

Humans can easily learn new concepts from just a single exemplar, mainly due to their remarkable ability to imagine or hallucinate what the unseen exemplar may look like in different settings. Incorporating such an ability to hallucinate diverse new samples of the tracked instance can help the trackers alleviate the over-fitting problem in the low-data tracking regime. To achieve this, we propose an effective adversarial approach, denoted as adversarial "hallucinator" (AH), for robust visual tracking. The proposed AH is designed to firstly learn transferable non-linear deformations between a pair of same-identity instances, and then apply these deformations to an unseen tracked instance in order to generate diverse positive training samples. By incorporating AH into an online tracking-by-detection framework, we propose the hallucinated adversarial tracker (HAT), which jointly optimizes AH with an online classifier (e.g., MDNet) in an end-to-end manner. In addition, a novel selective deformation transfer (SDT) method is presented to better select the deformations which are more suitable for transfer. Extensive experiments on 3 popular benchmarks demonstrate that our HAT achieves the state-of-the-art performance.