Background Matters: A Cross-view Bidirectional Modeling Framework for Semi-supervised Medical Image Segmentation

Luyang Cao, Jianwei Li, Yinghuan Shi*

Code

• Current semi-supervised medical image segmentation methods primarily focus on foreground modeling while overlooking the potential of background modeling. We demonstrate that reliable background modeling predictions enhance the confidence of foreground modeling. Building on this insight, we propose a Cross-view Bidirectional Modeling (CVBM) framework for semi-supervised medical image segmentation—the first work to integrate background modeling as an auxiliary view to enhance foreground modeling.

Towards Perfection: Building Inter-component Mutual Correction for Retinex-based Low-light Image Enhancement

Luyang Cao, Han Xu, Jian Zhang, Lei Qi, Jiayi Ma, Yinghuan Shi*, Yang Gao

Code

• Current Retinex-based low-light image enhancement methods typically assume that illumination and reflectance components can be perfectly decomposed and independently enhanced, while overlooking the inter-component residuals caused by imperfect decomposition. We demonstrate that these residuals lead to inaccurate component enhancement, resulting in color distortion and detail loss in the reconstructed images. Building on this insight, we propose an Inter-correction Retinex model (IRetinex)—the first framework to formally identify inter-component residuals and mitigate their effects through mutual correction during both the decomposition and enhancement stages.

Training Medical Volumetric Super-Resolution Model in ONLY One Epoch

Luyang Cao, Jian Zhang, Zekun Li, Lei Qi, Yinghuan Shi*

Code

• Existing medical volumetric super-resolution methods typically require hundreds or even thousands of training epochs, limiting their rapid adaptation in multi-center studies and medical emergencies. We identify slow shape reconstruction as a critical bottleneck and demonstrate that the strong shape representation capability of the Segment Anything Model can effectively guide early-stage reconstruction. Building on this insight, we propose ESAone, the first framework to explore accelerated training for medical volumetric super-resolution, achieving competitive high-resolution reconstruction in only one epoch through adaptive shape feature extraction, cross-axis frequency reconstruction, and shape-prior distillation.