Medical image segmentation plays a crucial role in clinical diagnosis and treatment planning. In recent years, generative methods such as diffusion models have shown strong potential for medical image segmentation because they can model complex data distributions and capture aleatoric uncertainty. However, these methods usually require hundreds of iterative sampling steps during inference, which results in high computational cost and limits their practical use in clinical environments. In addition, when generative segmentation models are applied to three dimensional medical image segmentation tasks, they often face severe GPU memory limitations. To address these issues, we propose a conditional Flow Matching (CFM) based framework for efficient 3D medical image segmentation. By adopting a straight probability path derived from optimal transport, the proposed method can generate high quality segmentation masks with only a single ODE integration step during inference. To alleviate the memory limitation caused by volumetric data and the loss of spatial location information in conventional patch training, we introduce a coordinate-aware patch training strategy, in which normalized 3D Cartesian coordinates are explicitly concatenated with local image patches, enabling the model to retain global anatomical positional awareness under constrained GPU memory. Experiments on the BraTS2020 brain tumor dataset and a private prostate MRI dataset demonstrate that the proposed method maintains competitive segmentation accuracy while achieving an inference time of only 1.14 seconds per volume, offering a clear efficiency advantage over existing generative segmentation methods and providing a viable solution for efficient 3D medical image segmentation.
Y. Yilihamu, Jian Xue, Chuang Jia et al.· International Conference on...· 0 citations
Visual grounding with multimodal large language models is commonly formulated as autoregressive coordinate generation, where a model outputs bounding-box coordinates as text given an image and a referring-expression prompt. While this interface is simple and compatible with instruction following, it introduces a mismatch between training and evaluation: training optimizes token-level likelihood over coordinate strings, whereas grounding quality is measured by geometric overlap. We propose IoU-PD, an IoU-aware privileged distillation method for coordinate-generating multimodal large language models. IoU-PD uses ground-truth boxes not only as coordinate targets, but also as privileged training-time guidance. During training, the student receives the original image and prompt, while a frozen teacher receives a box-marked image and an augmented prompt that indicates the marked region. The student is trained with a supervised fine-tuning anchor and a privileged distillation loss whose token weights reflect both geometric importance and teacher reliability. At inference time, IoU-PD requires no box overlay, privileged hint, teacher branch, or additional prediction module. Experiments on standard referring-expression grounding benchmarks show consistent region-level improvements over strong coordinate-generating baselines, demonstrating that ground-truth boxes can provide useful privileged guidance beyond serving as coordinate labels. Project page: https://xyzzzh.github.io/IoU-PD/
Generative Vision-Language Models (VLMs) commonly treat bounding-box coordinates as independent output symbols, leaving numerical order and axis semantics implicit. We identify this representation as an important source of error in visual grounding. Hi-Token encodes each coordinate with axis-specific tokens for the hundreds, tens, and ones digits, which adds coarse-to-fine structure and increases token reuse while retaining the existing VLM architecture. Hi-GAR complements this representation with a geometry-based reward for Group Relative Policy Optimization (GRPO), using box overlap and coordinate accuracy at multiple scales. Controlled comparisons under matched training conditions show that Hi-Token improves localization throughout the evaluated IoU range. Hi-GAR further reduces low-overlap predictions and is used only during training. Experiments on three VLM backbones and the RefCOCO family show consistent gains across models and benchmarks. Hi-R1 achieves higher values than strong specialist baselines on most reported metrics. Analyses of token frequency, digit boundaries, object scale, and IoU distributions explain the effects of coordinate representation and reward training. The results show that structured coordinate generation provides an effective approach to generative visual grounding. Project page: https://xyzzzh.github.io/Hi-Token/