Untrained Fixed Point Network

无训练不动点网络

• Oriented towards Untrained framework: Representative approach is Deep Image Prior, where the network doesn't require pre-training. Instead, it only trains on a single test sample during inference, and the network output becomes the test result after training.
• 面向无训练框架：代表方法是Deep Image Prior，网络不需要预训练，仅在推理时对单个测试样本进行训练，网络输出即为测试结果。
• First to combine fixed point theory with Untrained framework, with mathematical proof of convergence
• 首次将不动点理论与无训练框架结合，并提供收敛性的数学证明
• Incorporated prior information to design a physics-driven network architecture within the proposed framework
• 引入先验信息，在所提框架内设计物理驱动的网络架构
• Related work submitted to IEEE TIP (Q1 TOP, IF: 13.7), Major Revision
• 相关工作已投稿IEEE TIP（Q1 TOP，IF: 13.7），大修

Hypergraph-Inspired Linear Attention + Decoder-only Image Reconstruction Network

超图启发的线性注意力 + Decoder-only图像重建网络

• For MRI video reconstruction, data dimensions are large. Self-attention's quadratic complexity makes it difficult to directly apply in MRI video reconstruction frameworks (typically cascading dozens of networks)
• MRI视频重建中数据维度较大，自注意力的二次复杂度使其难以直接应用于MRI视频重建框架（通常级联数十个网络）
• Combined self-attention with graph theory, using hypergraph to restructure self-attention mechanism, transforming it into "aggregate features to hyperedges, then distribute features to nodes" operation, achieving linear complexity
• 将自注意力与图论结合，使用超图重构自注意力机制，将其转化为"聚合特征到超边，再分发特征到节点"的操作，实现线性复杂度
• For image restoration, current approaches directly apply U-Net from image segmentation
• 对于图像恢复任务，现有方法都是直接应用图像分割中的U-Net，但是没考虑U-Net到底适不适合
• Designed a Decoder-only network without U-Net's downsample-upsample approach, directly restoring from low to high resolution progressively, which is intuitive and suitable for image restoration tasks
• 设计了Decoder-only网络，摒弃U-Net的下采样-上采样方式，直接从低分辨率逐步恢复到高分辨率，直观且适合图像恢复任务
• Related work submitted to Medical Image Analysis (Q1 TOP)
• 相关工作已投稿Medical Image Analysis（Q1 TOP）

Tensor Low-Rank and Sparse Constrained Deep Unfolding Network

张量低秩稀疏约束深度展开网络

• For MRI video reconstruction, data is 3D (2D+T). Existing models mainly use matrices to describe redundant information within videos, which is insufficient
• MRI视频重建中，数据是3D的（2D+时间）。现有模型主要用矩阵描述视频内的冗余信息，这是不够的
• From tensor perspective, based on tensor t-SVD decomposition, mathematically derived and proved a new unitary transform-based t-SVD decomposition method, designing a tensor low-rank prior
• 从张量角度出发，基于张量t-SVD分解，数学推导并证明了一种新的基于酉变换的t-SVD分解方法，设计了张量低秩先验
• Combined tensor low-rank and sparse priors, derived iterative optimization algorithm, implemented with deep learning to build deep unfolding network, where low-rank and sparse transform domains are adaptively learned using CNN
• 结合张量低秩和稀疏先验，推导迭代优化算法，用深度学习实现构建深度展开网络，其中低秩和稀疏变换域由CNN自适应学习
• Related work published in CIBM (Q2 TOP at submission)
• 相关工作发表于CIBM（投稿时Q2 TOP）

Differentiable SVD Algorithm Research

可微分SVD算法研究

• In 2024, in PyTorch, if forward propagation contains SVD, training becomes extremely unstable
• 2024年，在PyTorch中，如果前向传播包含SVD，训练会变得极其不稳定
• The reason is: SVD derivative doesn't exist when there are two identical singular values
• 原因是：当存在两个相同的奇异值时，SVD导数不存在
• Deeply studied SVD and its derivative derivation, combined with matrix generalized inverse, defined a differentiable SVD, and implemented its forward and backward propagation in PyTorch, defining a new SVD operator
• 深入研究SVD及其导数推导，结合矩阵广义逆，定义了可微分SVD，并在PyTorch中实现其前向和反向传播，定义了新的SVD算子

CMRxRecon2025 Challenge - 8th Place

CMRxRecon2025挑战赛 - 第8名

Team: Hit.imip

队名：Hit.imip

• CMRxRecon2025 is a MICCAI 2025 challenge focused on cardiac MRI reconstruction, aiming to build foundation models for multi-center, multi-vendor, and multiple disease scenarios
• CMRxRecon2025是MICCAI 2025的挑战赛，聚焦心脏MRI重建，旨在构建多中心、多厂商、多疾病场景的基础模型
• Challenge involved 5+ centers, 10+ scanners, and various cardiovascular diseases including hypertrophic cardiomyopathy and myocardial infarction
• 挑战赛涉及5+中心、10+扫描仪，以及多种心血管疾病包括肥厚型心肌病和心肌梗死
• Our team achieved 8th place by fusing PromptMR+ with Swin Transformer architecture
• 我们通过融合PromptMR+与Swin Transformer架构获得第八名
• Trained with only 2 NVIDIA A6000 GPUs, potential for better results with more computational resources (for example, 8 GPUs)
• 仅使用2张NVIDIA A6000 GPU训练，如果有更多计算资源（例如8卡训练），能取得更好成绩

WPS2Latex (Vibe Coding project, powered by GitHub Copilot and Hermes Agent)

WPS2Latex (Vibe Coding项目，由GitHub Copilot和Hermes Agent驱动)

• A WPS JS macro tool similar to Excel2LaTeX, enabling one-click export of selected spreadsheet tables to LaTeX tabular code
• 类似Excel2LaTeX的WPS JS宏工具，支持一键将选中的表格导出为LaTeX tabular代码
• Supports bold/italic text, column alignment (l/c/r), multirow, multicolumn, and booktabs rules
• 支持加粗/斜体文本、列对齐（l/c/r）、合并行、合并列和booktabs规则
• Provides .xlam macro file that can be loaded directly into WPS Office
• 提供.xlam宏文件，可直接加载到WPS Office中
• MIT License, tested on WPS Office 2026 (Linux)
• MIT许可证，在WPS Office 2026（Linux）上测试