27 May 2025 in Machine Intelligence Research
Efficiently capturing multi-scale local information and building long-range dependencies among pixels are essential for medical image segmentation because of the various sizes and shapes of the lesion regions or organs. In this paper, we propose the multi-scale cross-axis attention (MCA) mechanism to address these challenges through enhanced axial attention. To address the issues of insufficient learning of positional bias and limited long-distance interaction in axial attention caused by the small dataset, we propose using a dual cross-attention mechanism instead of axial attention to enhance global information capture. Meanwhile, to compensate for the lack of explicit attention to local information in axial attention, we use multiple convolutions of strip-shaped kernels with different kernel sizes in each axial attention path, which improves the efficiency of MCA in local information encoding. By integrating MCA into the multi-scale cross-axis attention network (MSCAN) backbone, we develop our network architecture, termed MCANet. With merely 4 M+ parameters, MCANet outperforms previous heavyweight approaches (e.g., swin transformer-based methods) across four challenging tasks: skin lesion segmentation, nuclei segmentation, abdominal multi-organ segmentation, and polyp segmentation. The code is available at https://github.com/haoshao-nku/medical_seg .
10 June 2026 in IEEE Transactions on Circuits and Systems for Video Technology
Visual token compression is critical for Large Vision-Language Models (LVLMs) to efficiently process high-resolution inputs. Existing methods that typically adopt fixed compression ratios cannot adapt to scenes of varying complexity, often causing imprecise pruning that discards informative visual tokens and results in degraded model performance. To address this issue, we introduce a dynamic pruning framework, GlimpsePrune, inspired by human cognition. It takes a data-driven “glimpse” and prunes irrelevant visual tokens in a single forward pass before answer generation. This approach prunes 92.4% of visual tokens while on average nearly preserving the original performance on free-form VQA tasks, causing end-to-end latency and peak memory to reduce to 66% and 67% of the baseline, respectively. By integrating our pruning mechanism into the reinforcement learning fine-tuning stage, the time and memory costs of fine-tuning are also reduced with negligible performance loss. Code is available at https: //github.com/HVision-NKU/GlimpsePrune.