UniSpine-GS An Efficient Physics-Aware Gaussian Framework for Cross-Modality Multi-view Spine Image Synthesis

Jul 16, 2026

Upload by Violet

Spinal disease diagnosis relies on 3D imaging, yet accurate assessment is constrained by costly hardware and by physical differences across imaging modalities that reduce model generalizability. UniSpine-GS introduces an efficient, physics-aware Gaussian framework for novel-view projection rendering in multi-view spine imaging using a 3D-aware representation. Rather than explicit reconstruction, it learns geometry-aware Gaussians for anatomical consistency. SPWM improves boundary fidelity and local details, evaluated on CTSpine3D and FeSpine3D with code released publicly.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Abstract

Keywords

1 Introduction

Clinical motivation and challenges

Related work in neural rendering and Gaussian representations

Overview of UniSpine-GS and contributions

arXiv :2607 .04923v 1 [ cs .CV] 6 Jul 2026 UniSpine-GS: An Eﬃcient Physics-Aware Gaussian Framework for Cross-Modality Multi-view Spine Image Synthesis Qiuhua Chen 1 , Changning Yu 1 , Na Huang2 , Chao Sun 1 ,3(􀀀), and Bo Du 1 ,3 ,4 ,5 1 School of Computer Science, Wuhan University, Wuhan, China 2 The Department of Ultrasound of Renmin Hospital East Branch of Wuhan University, Wuhan, China 3 Institute of Artiﬁcial Intelligence, Wuhan University, Wuhan, China 4 National Engineering Research Center for Multimedia Software, Wuhan University, Wuhan, China 5 Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, Wuhan, China [chaosun@whu.edu.cn](chaosun@whu.edu.cn) Abstract. The diagnosis of spinal diseases is often assisted by 3D imaging techniques in clinical practice. However, precise 3D spinal assessment is limited by the high costs of 3D imaging hardware and the challenges posed by the physical diﬀerences between imaging modalities, which hinder the generalizability of models. To address these issues, we propose UniSpine-GS, an eﬃcient, physics-aware Gaussian framework designed for novel-view projection rendering in multi-view spine imaging via a 3D-aware representation. Instead of performing explicit 3D reconstruction, our approach learns a geometry-aware Gaussian representation that ensures anatomical consistency across diﬀerent views. We introduce SPWM, a structure-guided loss reweighting strategy to improve boundary ﬁdelity and local details. We evaluate our method on the CTSpine3Ddataset and a newly constructed 3D fetal ultrasound dataset, FeSpine3D. Our results demonstrate that UniSpine-GS signiﬁcantly outperforms existing methods across all metrics, oﬀering a practical and cost-eﬀective solution for uniﬁed multi-view medical imaging. Our code is publicly available at [https://github.com/orangeisland66/UniSpine-GS](https://github.com/orangeisland66/UniSpine-GS). Keywords: Multi-view Spine Imaging · Cross-Modality 3D Synthesis · X-Gaussian Representation · Medical Image Synthesis 1 Introduction Spinal disorders such as scoliosis, vertebral degeneration, and spondylolisthesis require an accurate assessment of spinal anatomy in three dimensions (3D) for reliable diagnosis and treatment planning [13] . In clinical practice, X-ray imaging is widely used due to its low cost, low dose, and broad accessibility, 2 Q. Chen, C. Yu et al. while ultrasound is highly valuable for bedside examinations and obstetric scenarios because it is radiation-free, portable, and real-time. However, accurate 3D spinal assessment remains challenging. High-quality 3D imaging and dedicated acquisition systems are expensive and not always available. Moreover, large differences in imaging mechanisms and visual appearance across modalities make it diﬃcult for models to maintain structural consistency and generalization in multi-modality and multi-view settings [11, 26] . These limitations motivate the need for cost-eﬀective methods that enable cross-view and cross-modality spine image synthesis with reliable structural representations. Recent advances in neural ﬁelds and diﬀerentiable rendering have enabled learning 3D representations from limited views for novel-view synthesis. NeRF [16] achieves high ﬁdelity but is often computationally expensive, motivating accelerated neural-ﬁeld variants based on eﬃcient encodings and anti-aliasing designs [2, 17] . In contrast, 3D Gaussian Splatting (3DGS) uses explicit Gaussian primitives with highly parallel rasterization to substantially improve training and rendering eﬃciency [1, 12] . For projection imaging, radiative Gaussian methods integrate radiative forward models with Gaussian representations [4] and have been extended to rectiﬁed tomographic reconstruction and continuous-time 4D (dynamic) tomographic reconstruction [21, 24], while structure-aware priors improve stability under sparse-view conditions [5] . Beyond static synthesis,

FAQs

What problem does UniSpine-GS address in multi-view spine imaging?

It targets the difficulty of accurate 3D spinal assessment caused by expensive 3D hardware and modality-dependent physical differences that limit structural consistency and generalization across views and modalities.

How does UniSpine-GS generate novel views without explicit 3D reconstruction?

It learns a geometry-aware Gaussian representation and uses a unified forward operator for geometry-consistent novel-view projection rendering.

What role does SPWM play in the method?

SPWM is a structure-guided loss reweighting strategy that emphasizes vertebral boundaries and texture-informative regions, improving boundary fidelity and local details under sparse-view supervision with warm-up scheduling.

Jul 16, 2026

Upload by Violet