Qian Huang

Generative transport models learn to move samples from one distribution to another, for example from a patient's magnetic resonance image to a computed tomography (CT) scan or from noise to an image. Recent ones, such as flow and bridge models, are used widely in medical imaging, and most of these studies report faster sampling or higher image scores than diffusion. In medicine these measures are incomplete; a synthetic CT scan, for example, must also keep the patient's anatomy and support dose calculation. We therefore review 383 studies of flow and bridge models published from 2023 to 2026, identified by a systematic search, with comparison samples of 55 diffusion and 17 normalizing-flow studies, and compare them across clinical tasks on model design and evaluation endpoints. We find that a model's medical value is shaped by the path the patient's information takes to the output, more than by whether the transport is a diffusion, flow or bridge variant. We present these models as settings of one transport viewpoint with six design axes. We then review four task families, reconstruction, translation, generation and analysis, and distil fifteen observations on the information a model receives and on how well its metrics and comparisons support its claims. For each clinical use we identify the endpoint it requires, which conventional metrics do not replace, together with the datasets and sampling costs a fair comparison needs, and give a procedure for choosing and testing a model. A continually updated reading list is maintained at https://github.com/chqwer2/Awesome-Flow-Based-Medical-Imaging.

In the realm of deep learning applied to computer vision, the extraction of features across multiple spatial scales stands as a pivotal requirement for achieving superior performance in tasks such as object detection, semantic segmentation, and image classification. Objects in natural scenes frequently exhibit significant variations in size, resolution, and contextual positioning, which traditional convolutional neural networks (CNNs) struggle to handle effectively due to their dependence on fixed kernel sizes and conventional pooling mechanisms. These limitations often lead to substantial information loss, reduced accuracy in recognizing fine-grained details, and inadequate capture of broader environmental contexts, ultimately hindering the model's generalization capabilities in real-world scenarios. To mitigate these challenges, this research introduces an innovative hybrid architecture that strategically fuses Spatial Pyramid Pooling (SPP) and Inception modules, leveraging their complementary strengths to create a more resilient and versatile framework for multiscale feature processing. SPP contributes essential spatial invariance by employing a hierarchical pooling strategy that aggregates features at various resolutions, allowing the model to process inputs of arbitrary dimensions without the need for distortion-inducing resizing or cropping. This mechanism generates fixed-length feature vectors that preserve critical spatial hierarchies, making it particularly advantageous for applications involving variable image sizes, such as satellite imagery analysis or medical diagnostics where scan resolutions can differ widely. Meanwhile, Inception modules enhance intra-layer feature diversity through parallel convolutional paths with differing kernel sizes, typically 1×1 for dimensionality reduction, 3×3 for mid-level patterns, and 5×5 for expansive contextual understanding, enabling the simultaneous capture of local intricacies and global structures within a computationally efficient design. By sequencing Inception's multi-kernel operations followed by SPP's pyramid aggregation, the hybrid model not only diversifies initial feature maps but also ensures scale-agnostic representations, addressing the core deficiencies of standard CNNs and promoting better integration of fine details with high-level semantics. To validate the effectiveness of this approach, comprehensive evaluations were conducted on well-established benchmarks: the CIFAR-10 dataset for image classification, which consists of 60,000 low-resolution color images across 10 categories and serves as a testbed for generalization under constrained conditions, and the Pascal VOC 2007 dataset for object detection, featuring 9,963 images with annotated bounding boxes for 20 classes in complex, multi-object scenes that challenge scale and occlusion handling. The proposed SPP-Inception hybrid achieved a classification accuracy of 94.3% on CIFAR-10 and a mean Average Precision (mAP) of 78.5% on Pascal VOC, representing improvements of 2.6% and 5.0%, respectively, over baseline architectures including standard CNNs, Inception-only models, and SPP-only configurations. These gains highlight the model's enhanced robustness in multi-scale environments, where it outperforms isolated components by better balancing local precision and global invariance. Furthermore, the hybrid maintains a favorable computational profile, with parameter counts and floating-point operations (FLOPs) only modestly higher than baselines, ensuring practicality for deployment in resource-limited settings such as mobile devices or edge computing platforms. While these results do not surpass absolute state-of-the-art figures, such as accuracies exceeding 99% on CIFAR-10 achieved by advanced transformer-based models like Vision Transformers (ViTs) or EfficientNet variants, they demonstrate competitive performance for a custom CNN hybrid, particularly in scenarios prioritizing interpretability and efficiency over sheer scale. The findings underscore the potential of this integrated design to advance deep learning architectures, offering a scalable solution for intricate visual recognition challenges. Future extensions could incorporate attention mechanisms or transformer elements to further refine adaptability, positioning the model as a foundational tool in fields ranging from autonomous systems and healthcare to environmental monitoring and beyond.

Skeleton-based human action recognition has emerged as a prominent research focus in computer vision, with significant progress achieved in recent years. However, existing methods often suffer substantial performance degradation under real-world data constraints, such as body occlusion, missing frames, and noise. These limitations critically undermine the robustness of related techniques in practical applications. To address these challenges, we propose a Spatial Temporal Self-compensating Graph Convolutional Network (STSc-GCN), which skillfully utilizes the systematic and regular nature of human movement to mitigate performance degradation caused by data constraints through a data self-compensation mechanism. Specifically, STSc-GCN comprises two key modules: 1) Collaborative Motion Spatial Compensation (CMSC). This module designs multiple distinct topological relationships, primarily including Walk-probability Generality Topology and Self-organizing Particularity Topology, respectively, to deeply explore the universal and personalized collaborative relationships between human joints. These relationships help compensate for the lack of information caused by spatial data constraints. 2) Meta-action Sharpening Temporal Compensation (MSTC). This module introduces a novel motion sharpening mechanism that enhances key dynamic information within the meta-action sequences through cross-attention technology, thereby improving model adaptability to missing-frame scenarios. STSc-GCN achieves state-of-the-art performance on four constrained datasets and shows superior results on three widely used standard datasets, confirming its effectiveness in both constrained and general scenarios. Code will be available at https://github.com/XingLi1012/STSc-GCN.git.

Human action recognition is a key research area in computer vision, where accurate recognition relies on effective modeling of both global and local spatiotemporal information. However, existing GCN-based methods often overemphasize the local topological connectivity of human skeletons. Moreover, their temporal modules fail to fully capture the evolution of action sequences, leading to critical instantaneous information being obscured by global representations. To address these problems, we propose an integrated framework termed MADS-GCN. In the spatial modeling stage, we introduce two parallel streams: the Physical Stream uses the adjacency matrix to constrain convolution and capture global structural patterns, while the Topological Stream leverages spatial attention to assign adaptive weights to joints, preserving discriminative local adaptive features. For temporal modeling, a channel-temporal attention mechanism is applied to adaptively refine feature maps, followed by a bidirectional GRU to capture multi-scale temporal patterns. Extensive experiments on NTU RGB+D60, Northwestern-UCLA, and our custom DanceBasic-Set demonstrate the effectiveness of MADS-GCN and indicate its applicability to dance action recognition scenarios.

Neural video compression (NVC) has emerged as a promising paradigm for improving rate-distortion performance. However, existing neural video codecs predominantly rely on convolutional neural networks (CNNs) with limited local receptive fields to generate the latent representations, often neglecting global–local spatial correlations. This leads to suboptimal feature modeling and redundancy in the latent space. To address this limitation, we propose a novel Dual-Scale Transformer (DST) block specifically tailored for NVC, which effectively enhances coding efficiency. The DST block incorporates a Global–Local (Shifted) Window-based Self-Attention (GL(S)WSA) mechanism to jointly capture global structure information and local texture details. Moreover, we design a Cross-Gated Feed-Forward Network (CGFFN) to adaptively modulate complementary components, producing more compact and expressive latent representations. Furthermore, to overcome the drawbacks of traditional asynchronous training and further boost rate-distortion performance, we introduce a Variable Bitrate Synchronization (VBRS) strategy that leverages multi-GPU parallel training, with each GPU dedicated to a specific bitrate and synchronized via gradient backpropagation for joint optimization. Experimental results demonstrate that our proposed method achieves the higher coding performance compared to the previous state-of-the-art (SOTA) methods and significantly outperforms H.266/VVC (VTM-13.2) under various low delay B (LDB) coding configurations.