Medical Image Volumetric Arbitrary Scale Super-Resolution Via Learnable Adaptive Upsampling and Weight Dynamic Balancing

Volumetric Computed Tomography (CT and Magnetic Resonance Imaging (MRI) are crucial in clinical diagnosis. However, practical limitations often result in anisotropic volumes with high in-plane but low through plane resolution. Volumetric SR therefore focuses on recovering missing through-plane details. Stronger anatomical variation along the through-plane makes direct adaptation of 2D restoration components prone to inefficient context aggregation and limited robustness in volumetric settings. In addition, although recent arbitrary-scale SR methods enable flexible scaling, they often face a trade-off between scale flexibility and parameter efficiency. To overcome these challenges, we propose a Learnable Adaptive Upsampling-based Volumetric SR network (LAUVSR). The first key innovation is a volumetric-anisotropy-driven upsampling core, which integrates a learnable weight generation module employing in-plane weight sharing to capture fine details and through-plane independent generation to model anatomical variations, augmented by a volume specific criss-cross attention mechanism through mapping attention to in-plane and through-plane interactions, ensuring both high-fidelity and parameter-efficient SR at arbitrary integer and decimal query scales. Furthermore, we in corporate a reinforcement-learning-guided weight dynamic balancing strategy to improve training robustness across datasets, modalities, and target scales. Extensive image quality assessments were conducted on four datasets across two modalities (CT and MRI) using synthetically downsampled LR-HR pairs. These assessments demonstrate the effectiveness of our proposed method in achieving high-quality SR results across a dense range of integer and decimal upsampling scales.Additionally, evaluations at unseen upsampling scales and downstream task analyses further support its robustness and practical potential.