Overview
Initiative type
Model of Care
Status
Deliver
Published
June 2026
Topic
Summary
This study evaluates whether artificial intelligence can generate synthetic 3D MRI suitable for accurate segmentation of thoracolumbar vertebrae in adolescent idiopathic scoliosis (AIS)..
Dates: 2026
Implementation sites: Children's Health Queensland in collaboration with QUT
This project was presented as a Poster at CEQ Showcase 2026 (PDF 352KB).
Aim
To implement an AI-driven, MRI-only 3D spine reconstruction tool that improves clinical decision-making and supports radiation-free, more patient-centred surgical planning for AIS.
Outcomes
- AI-generated 'synthetic MRI' accurately reproduced thoracolumbar bone detail without paired thoracolumbar MRI information.
- Improved segmentation accuracy (Dice score 87.34% vs 86.1%) with the synthetic MRI dataset compared to thoracic-MRI-only model.
- Achieved full thoracolumbar 3D spine reconstruction without CT scans, supporting scalable and radiation-free 3D spinal evaluation for AIS.
- Provides clinicians with clearer, more accessible 3D models for planning AIS surgery, including in regional and rural centres.
Background
Magnetic resonance imaging (MRI) is increasingly preferred to X-Rays and CT scans for assessing the spine in children and adolescents because it provides detailed views of soft tissues and vertebrae without exposing patients to any ionising radiation. This makes MRI especially valuable in adolescent idiopathic scoliosis (AIS), where paediatric patients often require repeated imaging throughout many years of growth, bracing, surgery planning, and post-op follow-up. Compared with computed tomography (CT), MRI avoids cumulative radiation dose while offering excellent visualisation of the spinal cord, discs, and surrounding soft tissues, making it a safer, radiation-free choice for long-term spine care.
However, despite these advantages, routinely acquired MRI sequences at most centres focus on detecting neural axis anomalies and lack the 3D resolution required for high fidelity reconstruction of osseous anatomy. While adequate for soft tissue and qualitative assessment, these standard clinical MRI capture insufficient detail and the limited resolution prevents accurate 3D vertebral reconstruction.
Surgeons and clinicians increasingly rely on 3D models of the spine to assess deformity, track progression, and plan corrective procedures. But a major barrier to providing them is the lack of labelled, high-resolution, full-spine 3D MRI datasets available. Creating such datasets from standard clinical MRI sequences requires manually outlining every vertebra in 3D, on each individual MRI slice, which is an extremely time-consuming and a specialised task. Because of this, large-scale labelled MRI datasets for paediatric scoliosis simply do not exist.
In contrast, historical full-spine CT datasets, often ordered in the past to assist with surgical planning have detailed vertebral labels of AIS patients' spines, and provide rich anatomical information. However, CT and MRI have fundamentally different appearances, contrast mechanisms, and noise patterns. AI models trained on CT generally do not transfer well to detecting bony margins on MRI, a problem known as cross-modality domain shift. This is particularly challenging in scoliosis, where capturing deformity-specific bone shapes is essential. As a result, existing CT-labelled datasets cannot be directly used to train MRI-based AI bony segmentation models.
This project addresses that challenge. We developed an AI framework that bridges the gap between CT and MRI by synthetically generating MRI-like images from accurate CT data. These 'pseudo-MRIs' preserve the accurate vertebral shapes from thoracolumbar scoliosis CT but look like MRI, allowing CT-derived labels to be reused for MRI-based AI model training when in reality, true thoracolumbar 3D MRI data are not currently available.
Our goal is to enable accurate, automated full-spine bony segmentation using 3D MRI alone at inference, while leveraging historical CT datasets only during AI model training. By doing so, we aim to support radiation-free, high-quality 3D full spine assessment for paediatric patients with AIS in the future, that may replace CT for the purposes of both surgical planning and surgical navigation.
Methods
To overcome the absence of comprehensive labelled thoracolumbar 3D MRI datasets, we first generated synthetic MRI-like volumes from historical low dose T1-L5 thoracolumbar CT scans that had been manually labelled (Ethics #ERM 25276). Although CT is no longer used in routine AIS assessment, these historic datasets offered detailed vertebral annotations spanning thoracic and lumbar regions. A generative adversarial network (GAN) was trained to transform these CT volumes into images that closely mimic the visual characteristics of MRI. During training, informed by an unpaired thoracic-only 3D MRI dataset (Ethics #ERM 24390), the GAN learned to reproduce MRI-like soft tissue contrast while preserving the exact bone morphology and deformity specific anatomy contained in the original CT of scoliosis patients. This step effectively bridged the appearance gap between CT and MRI, allowing us to retain the anatomical richness of CT while producing images compatible with MRI-based AI algorithm, and producing a full thoracolumbar spine synthetic MRI dataset.
The synthetic thoracolumbar pseudo MRI dataset was then combined with an existing 3D T1 weighted thoracic MRI dataset collected for a scoliosis progression research project. These research 3D MRIs, although limited to the thoracic spine, provided authentic MRI contrast and patient specific deformity patterns essential for model realism. The combined datasets therefore offered both complete spine bony anatomical coverage - derived from the pseudo MRI volumes - and true MRI signal characteristics from the thoracic 3D MRI research scans.
A U Net-based deep learning architecture was then trained on the combined datasets to perform bony segmentation across the entire thoracolumbar AIS spine. During training, the model learned to integrate information from both the synthetic thoracolumbar and thoracic 3D MRI domains, developing the capacity to infer and reconstruct the lumbar spine even when the raw input research MRI dataset contained only thoracic coverage. The model was optimised to maintain continuity between vertebral levels and to accurately capture scoliosis curve morphology, rotational vertebral deformities, and anatomical boundaries crucial for clinical and surgical use.
Model performance was evaluated on a hold out set of AIS 3D MRI scans that were not used during training. We assessed whether the model could reliably generate anatomically consistent thoracic spine segmentations and whether the inclusion of pseudo MRI data improved segmentation accuracy and generalisability. Dice similarity coefficient was used as the primary quantitative metric. Comparative evaluation demonstrated that AI segmented spine models trained with pseudo MRI achieved higher accuracy and produced smoother, more coherent reconstructions than those trained using the research thoracic 3D MRI data alone. Importantly, the final trained model can produce a full spine 3D bony reconstruction using 3D MRI scans as the only input at inference, supporting its intended role as a radiation free, clinically integrated digital tool for AIS assessment and surgical planning.
Discussion
Successful development of this AI framework depended on several key enablers in both the clinical and technical environment.
First, access to high quality 3D thoracic MRI scans from AIS patients from a prior research project provided a reliable foundation for AI model training and validation. This 3D thoracic MRI dataset reflected the capability of clinical MRI should a 3D sequence be ordered, ensuring that the system was built around practical and achievable workflows.
Second, the availability of a historical, fully labelled thoracolumbar CT databank was essential for generating pseudo MRI volumes that provided the anatomical coverage of the whole spine missing from the available research thoracic spine 3D MRI dataset. This cross-modality pairing created the conditions necessary for the AI model to learn anatomy beyond the thoracic region despite incomplete 3D MRI inputs. Equally important was the strong collaboration between specialist spine deformity clinicians, biomedical engineers, and AI engineers. Clinical oversight ensured that synthetic MRI data and bony segmentation outputs maintained anatomical fidelity, while technical expertise allowed complex generative and segmentation models to be implemented effectively.
Several lessons were learned during the course of the project. Most notably, synthetic MR image generation can meaningfully enhance MRI based AI workflows when high quality labelled MRI data are limited. The project demonstrated that pseudo MRI can increase anatomical completeness and segmentation continuity, even when the available 3D MRI dataset had captured only the thoracic spine. Another key lesson was the importance of validating AI outputs against clinical expectations rather than relying exclusively on numeric performance metrics. Anatomical plausibility, scoliosis spine morphology, and continuity across vertebral regions emerged as equally important indicators of AI model suitability for clinical deployment. The project also underscored the need for diverse imaging datasets: while CT-derived labels were valuable, their anatomical patterns may differ from those typically observed in AIS, highlighting the need for ongoing refinement. Limitations include the retrospective nature of the historical imaging datasets, variation in MRI protocols across centres, and the lack of full-spine ground truth 3D MRI for direct comparison. Although the model reconstructs thoracolumbar anatomy effectively, further evaluation will be required before it is used directly for clinical decision-making involving surgical correction magnitude or screw trajectory planning (surgical navigation).
Importantly, the framework developed here is inherently forward-compatible with future patient imaging improvements. If thoracolumbar 3D MRI becomes routinely acquired in AIS care - whether due to advancements in faster sequences, improved patient tolerance, or evolving clinical guidelines - the model will be well positioned to incorporate and benefit from these richer and more numerous inputs. In such a scenario, the AI system could transition from inferring the thoracolumbar region to directly segmenting it, further enhancing accuracy and clinical utility for deformity assessment and surgical planning.
One important potential use for 3D MRI-derived reconstructions is their integration into spinal navigation robots, which currently rely on CT-based bony anatomy to guide screw placement. Achieving bony segmentation on 3D MRI with sufficient accuracy for navigation would represent a major advance in reducing the risk of screw misplacement.
This project has clear applicability across Queensland Health. The workflow is inherently scalable across health facilities.
References
1. Goldman SN, Hui AT, Choi S, Mbamalu EK, Tirabady P, Eleswarapu AS, et al. Applications of artificial intelligence for adolescent idiopathic scoliosis: mapping the evidence. Spine Deform. 2024;12:1545–70.
2. Shcherbakova YM, Lafranca PPG, Foppen W, Ito K, Seevinck PR, Schlosser TPC. A multipurpose, adolescent idiopathic scoliosis-specific, short MRI protocol: A feasibility study in volunteers. Eur J Radiol. 2024;177:111542.
3. Silveira JT, S G, Kundapur PP. Automated lumbar spine segmentation in MRI using an enhanced U-Net with inception module and dual-output mechanism. Sci Rep. 2025.
4. Ji Y, Mei X, Tan R, Zhang W, Ma Y, Peng Y, et al. Three-dimensional automated segmentation of adolescent idiopathic scoliosis on computed tomography driven by deep learning: A retrospective study. Medicine (Baltimore). 2025;104(22):e42644.
5. Buijs RE, Cornelissen DM, Schlösser TPC. Quantification of thoracic volume and spinal length of pediatric scoliosis patients on chest MRI using a 3D U-Net segmentation. Healthcare (Basel). 2025;13(18):2327.
6. Nikbakhsh S, Naghashyar L, Valizadeh M, Amirani MC. Enhanced synthetic MRI generation from CT scans using CycleGAN with feature extraction. arXiv preprint. 2023.
7. Jin CB. DC2Anet: Generating lumbar spine MR images from CT scan data based on semi-supervised learning. GitHub / Appl Sci. 2019. Available from: https://github.com/ChengBinJin/SpineC2M
8. Kim KH, Lee EC, Yoon YD, Shin DW, Koo HW, Lee BJ. Translation of computed tomography images to T2-weighted magnetic resonance images of lumbar spine using generative adversarial networks. Sci Rep. 2025.
9. Vrettos K, Koltsakis E, Zibis AH, Karantanas AH, Klontzas ME. Generative adversarial networks for spine imaging: A critical review of current applications. Eur J Radiol. 2024;171:111313.
10. Hong KT, Cho Y, Kang CH, Ahn KS, Lee H, Kim J, et al. Lumbar spine computed tomography to magnetic resonance imaging synthesis using generative adversarial network: Visual Turing test. Diagnostics. 2022;12(2):530.
11. Li L. Development of 3D spatial measurement method based on computed tomography images of adolescent idiopathic scoliosis [dissertation]. Hong Kong: The Hong Kong Polytechnic University; 2025
Key contact
Dr Nathasha Naranpanawa
Postdoctoral Research Fellow - Science and Engineering
Children's Health Queensland in collaboration with QUT