Sparse-view CT reduces radiation exposure but renders image reconstruction severely ill-posed. Although deep learning has substantially improved reconstruction quality, most existing models are trained within specific anatomical domains and often degrade under cross-anatomy distribution shifts. To improve reconstruction generalization under limited target-domain supervision, we investigate a meta-learning framework for sparse-view CT reconstruction. However, directly applying Model-Agnostic Meta-Learning (MAML) to this task is challenging, because CT reconstruction is a high-dimensional inverse problem in which aggressive inner-loop updates can easily induce meta-overfitting and unstable optimization. To address this issue, we first introduce Stable Meta-Learning (StaML), which constructs anatomy-wise tasks and adopts conservative inner-loop updates to learn a more transferable initialization. Building on this, we further propose Adaptive Meta-Learning (AdaML), which dynamically adjusts per-step learning rates during meta-testing to better balance convergence capability and stability. Extensive experiments across multiple CT datasets and anatomical regions, under different k-shot settings, view numbers, and noise levels, demonstrate that AdaML consistently improves reconstruction quality over representative meta-learning and non-meta-learning baselines with favorable convergence. These results suggest that combining stable meta-initialization with adaptive test-time optimization is a promising direction for cross-anatomy sparse-view CT reconstruction. The source code is publicly available at https://github.com/Stth-er/AdaML.
Fully automated segmentation of medical images is essential for streamlining clinical workflows and improving diagnostic efficiency. However, building a single model that jointly handles multiple segmentation tasks remains challenging because different tasks may vary not only in target structures but also in heterogeneous data distributions. Existing learnable-prompt methods encode task-specific priors, but their generalization may still be constrained by cross-task interference when the tasks have substantially different data distributions. In this study, we propose Hybrid Prompt-based Segmentation, referred to as HybPromptSeg, which is a hybrid prompt-driven, encoder–decoder-based network that unifies multiple medical image segmentation tasks within a single shared model. Concretely, we introduce a Unified Hybrid Prompt (UHP) mechanism that integrates spatial-domain and Fourier-amplitude representations within a shared learnable prompt to model task- and modality-related variations. We further design a Task-conditioned Prompt Generation (TPG) module that generates discriminative task-conditioned prompts to guide the decoder for different segmentation tasks. To evaluate HybPromptSeg, we construct two benchmarks: a single-modality multi-task (SMMT) benchmark for evaluating task discrimination under shared-modality conditions, and a multi-modality multi-task (MMMT) benchmark for evaluating robustness under cross-modality variations. Experimental results demonstrate that HybPromptSeg achieves the best performance among the evaluated unified multi-task models on both benchmarks. The code is publicly available at https://github.com/nnnsjtu/HybSeg.
Accurately converting pixel measurements into absolute real-world dimensions remains a fundamental challenge in computer vision, limiting progress in applications such as biomedicine, forensics, nutritional analysis, and e-commerce. We introduce RulerNet, a deep learning framework that robustly infers scale in the wild by reformulating ruler reading as a unified keypoint detection problem and representing rulers with geometric progression parameters that compactly approximate the non-uniform spacing induced by perspective transformations. Unlike traditional methods that rely on handcrafted thresholds or rigid, ruler-specific pipelines, RulerNet directly localizes centimeter markings using a mark-visibility-based annotation and training strategy that remains valid under mark-preserving transformations, enabling strong generalization across diverse ruler types and imaging conditions while mitigating data scarcity. Additionally, we introduce a scalable synthetic data generation pipeline that combines graphics-based ruler creation with ControlNet-enhanced realism, significantly expanding training diversity and improving model performance. Extensive experiments on both our datasets and public benchmarks demonstrate that RulerNet achieves accurate, consistent, and efficient scale estimation under challenging real-world conditions. Integration into a medical analysis pipeline further demonstrates its practical utility for scale-aware measurement. These results suggest that RulerNet can serve as a generalizable measurement component and be readily integrated with other vision modules, such as segmentation, monocular depth estimation, and diagnostic workflows, for automated analysis in medical and other domains. An online demo is provided. The code and data are publicly released with the paper.
