Advancements and Clinical Implications of Deep Learning-Based Synthetic CT Generation from MRI for Spine Surgery: A Literature Review

Abstract

This study reviews the transformative impact of deep learning (DL) in generating synthetic computed tomography (sCT) images from magnetic resonance imaging (MRI) datasets, particularly in spine surgery. It explores how DL-driven sCT aims to enhance surgical planning, improve diagnostic capabilities, and potentially integrate with navigation and robotic systems, while also critically evaluating current methodologies, performance metrics, and challenges to widespread clinical adoption. The overarching goal is to reduce patient radiation exposure and streamline clinical workflows by providing CT-equivalent bone visualization from MRI data.

 

MissingLinkText

Keywords

artificial intelligence - deep learning - magnetic resonance imaging - neuronavigation - spine surgery - surgical planning - synthetic CT

Introduction

In spinal surgery, medical imaging is indispensable for accurate diagnosis, meticulous preoperative strategizing, and ensuring navigational accuracy during surgical procedures. Magnetic resonance imaging (MRI), with its superior soft-tissue contrast, and computed tomography (CT), offering unparalleled bone visualization, traditionally form the twin pillars of spinal imaging. However, the reliance on both modalities presents challenges; CT scans expose patients to ionizing radiation, which carries a lifetime risk of cancer, and acquiring images from two different modalities can interrupt clinical workflow, increase missed diagnosis or misdiagnosis rate, increase healthcare costs, and introduce complexities in image registration.[1] [2] [3]

A promising innovation, synthetic CT (sCT) generated from MRI data, primarily studied in radiotherapy, has emerged as a significant advancement in addressing these issues. The primary aim of MRI-derived sCT is to provide CT-like bone delineation from an MRI acquisition, thereby potentially eliminating the need for a separate CT scan. This could significantly reduce patient exposure to ionizing radiation, make the diagnostic and planning process more efficient, and reduce costs.[4] [5] [6] [7] Early attempts at sCT generation involved atlas-based methods or tissue segmentation approaches. Atlas-based methods rely on registering an MRI atlas (with a corresponding CT) to the patient's MRI and warping the atlas CT, but are highly dependent on registration accuracy and interindividual anatomical variations. Segmentation-based methods classify MRI voxels into different tissue types and assign predefined attenuation properties, often requiring multiple MRI sequences and facing challenges in classification accuracy.[8] [9]

Recently, deep learning (DL), particularly the application of Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), has revolutionized sCT generation. These advanced algorithms have demonstrated the capability to produce sCT images of such high fidelity that their quality, in terms of bone visualization and quantitative accuracy, can be comparable to that of actual CT scans, especially in the context of spine surgery.[4] [10] Thus, this review will explore the current DL methodologies for sCT generation in spine surgery, discuss their clinical applications and validation, analyze performance metrics, and consider the challenges and future directions for this technology.

 

MissingLinkText

Methods

A literature review was conducted for this study. The databases PubMed, IEEE Xplore, Medline, Ovid, and Google Scholar were searched for English and French studies published up to early 2024, focusing on DL-generated sCT in spinal applications. Keywords included “synthetic CT,” “MRI to CT conversion,” “spine surgery,” “deep learning,” “artificial intelligence,” and specific to spine or vertebral imaging. Original research articles were selected and scrutinized based on methodological robustness (e.g., DL architecture, dataset size, training strategy—paired vs. unpaired), image fidelity (qualitative assessment), validation rigor (quantitative metrics such as MAE, peak signal-to-noise ratio [PSNR], DSC, SSIM, geometric accuracy, and diagnostic performance), and reported or potential clinical impact. Although studies focusing primarily on robotic system development were not the central theme, the implications and integration of sCT with surgical navigation and robotics were considered key aspects of its clinical usefulness. From an initial pool of 50 articles found, 8 were included in this literature review (see [Table 1]).

Table 1 Summary of included study findings

Study	Study design	Population	Technical approach for deep learning algorithms	Performance indicators	Main finding
Cao et al, 2024[11]	Prospective study	19 patients aged 30–80 y. Low back pain with a diagnosis of lumbar disc herniation Exclusion: pregnant, metallic implants, and incomplete imaging examination. China	Conventional axial 2D T2W MR images with a slice thickness of 1 mm were used as input for the U-net model	Assessors: Two musculoskeletal radiologists with 15–20 y of experience. Assessment: Subjective image quality of the synthetic and conventional CT scans were assessed on a four-point scale. Sensitivity, specificity, and accuracy were used to compare the subjective assessment	The sensitivity, specificity, and accuracy of CT for the detection of lumbar disc herniation were as follows: conventional CT vs. sCT; sensitivity = 84–91% vs. 81%; specificity = 83–85% vs. 98–100%; and accuracy = 84–87% vs. 90–91% Synthetic and conventional CT images showed considerable agreement in the assessment of the lumbar spine. Thus, synthetic CT images can be useful in the diagnosis of lumbar disc herniation
Staartjes et al, 2021[13]	Proof-of-concept study	1 case was scheduled for robotic lumbar fusion surgery and 2 healthy volunteers. Europe	MRI images were acquired with a 1.5T field strength and a standard lumbar spine protocol including T1- and T2-weighted sequences and sagittal 3D T1W multi-gradient echo sequence for bone MRI reconstruction. Synthetic CT scans were generated from MRI inputs using a patch-based convolutional neural network, similar to U-Net	Visual inspection of the quality of the vertebral body height and the spinal canal diameter images of the sCT in comparison with T2W MRI and spinal CT with a mean absolute difference (MAD) statistical analysis	Comparative measurements of anterior and posterior vertebral body height and spinal central canal diameter were with a MAD of 0.26 (±0.24 mm). Qualitatively adequate sCT for surgical planning
van der Kolk et al, 2022[14]	Prospective study	25 patients aged over 50 y. Cervical radiculopathy Exclusion: Patients with osteosynthetic material in the cervical spine and known malignant bone tumors. The Netherlands	Standard MRI of the cervical spine was acquired with an Ingenia 1.5 T MRI system, supplemented with a generally applicable radiofrequency-spoiled T1-weighted multiple gradient-echo sequence were processed by the BoneMRI software to create the sCT volumes	Assessors: Three specialists (neurosurgeon, neuroradiologist, and musculoskeletal radiologist) with 10, 20, and 24 y of experience. Qualitative image assessment: Evaluation of anatomical structures in sCT and CT according to the European guidelines on quality criteria on a four-point scale. Quantitative image assessment: Geometrical accuracy [B.K, T.S., J.O.] and voxel-wise intensity-based comparisons [mean absolute error (MAE)/dice similarity coefficient (DSC)/structural similarity index (SSIM)]	Image quality assessment and geometrical accuracy show good to excellent agreement between sCT and CT. Voxel-wise comparisons showed a MAE of 80.05 (± 6.12) HU, a dice similarity coefficient (cortical bone) of 0.84 (± 0.04), and a structural similarity index of 0.86 (± 0.02); deep learning-based sCT was noninferior to conventional CT. sCT of the cervical spine produces adequate qualitative and quantitative image quality, providing optimal visualization of both soft and bony structures in a single MRI exam
Bajger et al, 2021[15]	Preliminary study of the retrospective study of Roberts et al, 2023	400 CT and 400 MRI of the lumbar spine performed between January 2015 and December 2019 at Flinders Medical Centre or the Royal Adelaide Hospital were used for the training of the deep learning model. For the preliminary study, 8 cases were used for evaluation. South Australia	CycleGAN on MRI and CT images with the resolution of 0.8 mm × 0.8 mm × 0.8 mm	Synthesized CT (sCT) images were quantitatively evaluated in comparison with the real CT images, with MAE and peak-signal-to-noise ratio (PSNR)	The average MAE was found to be 184 and the mean PSNR was found to be 19:60. The results are comparable to the literature
Roberts et al, 2023[16]	Retrospective study	From 400 unpaired CTs and 400 unpaired MRI, only thin-section CT acquisitions and sagittal T1-weighted MRI sequences were extracted from the lumbar spine images performed between January 2015 and December 2019 at Flinders Medical Centre of the Royal Adelaide Hospital. Data from 20 sCT from MRI were analyzed. South Australia	MRIs were mostly acquired in 1.5 tesla and CT scans from a mixture of Philips and Toshiba. 3D CycleGAN was used to enable MRI-to-CT and CT-to-MRI mappings	Assessors: Two neurosurgeons and two radiology registrars with 4, 7, 5, and 3 years of training experience. Assessment: The imaging evaluation was undertaken using ITK-SNAP for specific anatomical measurements (mid-vertebral body height, intervertebral disc height, and lumbar lordosis in the sagittal plane; and right and left pedicle in the axial plane)	Between the sCT and the real CT, the relative error of mid-vertebral body height was of 1 to 3% with S1 having a larger one of 8%. Intervertebral disc height was of 1 to -7% and a larger at -10% for L5–S1. Lumbar lordosis was of 10%. And right and left pedicles in axial plan were from a range of 34 to -19%. Measurements in sagittal plane were highly correlated, unlike axial plane measurements that were poorly correlated
Oulbacha and Kadoury, 2020[17]	Validation study	18 patients' MRI and CT images of the lumbar spine from the Public SpineWeb library. Canada	T2-weighted MRI data images were normalized and refined with an FC-ResNet to segment the vertebral bodies and pedicles from the sagittal slice. MRIs were standardized to have zero mean and unit variance and downscaled with CTs. The MR-CT synthesis network architecture used was automated and inspired by the 3D CycleGAN (pseudo-3D approach)	Dice scores were used to assess the vertebral morphology. MAE and normalized histogram intersection (NHI) were used as a voxel metric. Landmarks of the visible vertebrae were both segmentation and manual landmarks	The average dice was 0.83 ± 0.16, the MAE and the NHI were satisfying and suitable for the lumbar spine region and the landmark error between real and synthetic CT was 2.2 ± 1.4 mm. Thus, this CT synthesis network architecture demonstrates good potential for surgical guidance workflows
Jans et al, 2021[18]	Prospective study	30 patients aged 18–60 y, from an outpatient clinic, with inflammatory sacroiliac (SI) suspected back pain, for at least 3 mo. Exclusion: pregnant, metallic implants, contraindication to MRI, and incomplete imaging examination. Belgium	MRI, 3.0 T, with applied SI joint protocol and an additional axial 3D T1-weighted sequence, was used. On the same day, participants underwent an axial pelvic CT with a dual-energy CT protocol. Synthetic CT reconstruction was made by using commercially available software: BoneMRI Pelvic Region, version 1.1, based on the U-net architecture	Assessors: Two musculoskeletal radiologists of 16 and 5 y of experience. Assessment: Diagnostic confidence of every lesion (erosion, sclerosis, ankylosis) score was evaluated using a four-point scale, and sensitivity, specificity, and accuracy were used to compare the diagnostic performance between sCT and T1-weighted MRI	No differences were found in sensitivities for erosion or in specificities for sclerosis and ankylosis between sCT and T1-weighted MRI. All structural lesions could be more reliably depicted with sCT than with T1-weighted MRI, more for erosion and ankylosis. sCT had higher diagnostic accuracy and reliability than T1-weighted MRI, and with reliability comparable to that of CT, especially for erosion and ankylosis lesions
Morbée et al, 2021[19]	Prospective study	30 patients, from February 2019 to May 2020, who accepted to undergo an additional MRI of the lumbar spine on the same day as their CT of the lower lumbar spine. Exclusion: pregnant, metallic implants, and contraindication to MRI. Belgium	MRIs were obtained on a 3.0T using an axial 3D T1 multi-echo gradient echo FLASH sequence. On the same day, participants underwent an axial CT with a dual-energy CT protocol. sCT images were reconstructed in a coronal and sagittal direction with the research version of the BoneMRI software, version 1.3, based on the U-net architecture	Two radiologists of 8 and 5 years of experience, measured, independently, left and right pedicle width, spinal canal width, left and right neuroforamen length, anterior vertebral body height, posterior vertebral body height, superior vertebral body length, inferior vertebral body length, superior vertebral body width, inferior vertebral body width, maximal disc height, lumbar curvature, and spinous process length on sCT and CT	sCT is equivalent to CT in terms of geometrically accurate visualization of bony morphology of the lumbar spine; as of pedicle width, spinal canal width, vertebral body height, vertebral body width, vertebral body length, and spinous process length, except neuroforamen length and maximal disc height

 

MissingLinkText

Results

Deep Learning Algorithms in sCT

Within the DL paradigm, CNN frameworks such as U-Net and GANs, including CycleGANs, have become prevalent for image-to-image translation tasks, such as MR-to-sCT generation.

The U-Net, with its encoder–decoder structure and skip connections, is well-suited for image segmentation and reconstruction tasks. Cao et al[11] employed a U-Net model to generate sCT images from axial T2-weighted (T2W) MR images for diagnosing lumbar disc herniation in 19 patients. Their method aimed to provide CT-equivalent information directly from clinical MRI sequences. van Stralen et al,[12] in work foundational to Staartjes et al[13] and van der Kolk et al,[14] also utilized a U-Net-like patch-based CNN for their “BoneMRI” technique, trained on paired MRI-CT data using a specific 3D radiofrequency-spoiled T1W multiple gradient echo (T1W-MGE) MRI sequence as input.

Similarly, a significant advancement in medical image translation is the CycleGAN, which can learn mappings between two image domains (e.g., MRI and CT) using unpaired datasets, meaning that there is no need for perfectly aligned MRI and CT scans from the same individuals for training. This major advantage consists of a generator network that creates synthetic images and a discriminator network that distinguishes synthetic images from real ones, driving the generator to produce increasingly realistic outputs. Bajger et al[15] utilized a CycleGAN trained on a large clinical dataset of 800 patients with unpaired, unaligned T1W MR and CT images of the lumbar spine. Their model produced sCTs with an average mean absolute error (MAE) of 184 Hounsfield units (HU) (standard deviation [SD]: 24 HU) and a PSNR of 19.60 dB (SD: 1.20 dB) on a test set of eight paired cases. Roberts et al[16] employed a supervised full 3D CycleGAN trained on 400 unpaired CTs and 400 unpaired sagittal T1W MRIs to generate lumbar spine sCTs with a similar quality. To better incorporate volumetric information, Oulbacha and Kadoury[17] proposed a pseudo-3D CycleGAN for lumbar spine sCT synthesis from T2W MRI, using four neighboring 2D slices stacked together as input to approximate the 3D context without the high memory demands of full 3D convolutions. Their method, evaluated on 18 paired and aligned cases from the SpineWeb dataset, yielded an MAE ranging from 93 to 180 HU (average: 126 HU) and a dice similarity coefficient (DSC) of 0.83 for vertebrae segmentation on the sCTs. They noted that their pseudo-3D approach produced more visually coherent volumes than the standard 2D single-slice CycleGAN.

The choice of MRI sequence used as input for DL models varies. T1W sequences are common because of their generally good contrast for anatomical structures. The “BoneMRI” technique was the most used.[11] [13] [14] [15] [16] [17] [18] [19] Specifically, 3D T1W multiple gradient-echo sequence was used. Conversely, Cao et al[11] and Oulbacha and Kadoury[17] opted for T2W sequences.

Bahrami et al[2] have been pivotal in comparing different CNN architectures, showcasing their capability to synthesize sCT images with high fidelity from MR images. While their work and that of Li et al[7] demonstrate the generalizability of DL for this spine-specific task, accuracy and validation are still crucial.

 

MissingLinkText

Clinical Applications and Validation of sCT in Spine Surgery

A primary driver of sCT development is the elimination of radiation exposure associated with conventional CT scans. Staartjes et al[13] highlighted that in one of their cases, sCT could have prevented a spiral CT scan, thereby avoiding a volume CT dose index (CTDIvol) of 12.9 mGy. This benefit is particularly relevant for younger patients and those requiring multiple follow-up scans.

MRI-based sCT has also shown potential for enhancing the diagnostic capabilities of certain spinal pathologies. Cao et al[11] and Jans et al[18] demonstrated that sCT can achieve diagnostic performance comparable or even superior to conventional methods (CT or T1W MRI, respectively) for specific pathologies such as lumbar disc herniation and sacroiliac joint lesions.

Cao et al[11] investigated the diagnostic performance of U-Net-generated sCT from T2W MRI for lumbar disc herniation in 19 patients. They found that sCT had similar or better sensitivity, specificity, and accuracy compared with conventional CT, exceeding 80% for each, enhancing the visibility of lesions, concluding that sCT can be utilized for this diagnostic purpose.

Jans et al[18] evaluated a DL-based sCT (BoneMRI) derived from 3D T1MGE sequences for detecting structural lesions (erosions, sclerosis, ankylosis) in the sacroiliac joints of 30 patients suspected of having sacroiliitis, using conventional CT as the reference. Their findings indicated that sCT had significantly higher diagnostic accuracy than standard T1W MRI for detecting erosions (94 vs. 86%), sclerosis (97 vs. 81%), and ankylosis (92 vs. 84%). The sCT also improved the specificity for erosion detection and sensitivity for sclerosis and ankylosis detection compared with T1W MRI.

Also, van der Kolk et al[14] found their cervical sCT to be noninferior to conventional CT for overall bone image quality, artifacts, and visualization of intervertebral joints and neural foramina, though noninferiority was weaker for trabecular bone detail. They noted sCT's advantage in reducing metal artifacts from implants and improving visualization of the lower cervical spine compared with CT.

Another interesting part of sCT is its accuracy in geometric representation of the bony anatomy, which is crucial for preoperative planning, such as determining pedicle screw trajectories and sizes. Staartjes et al,[13] van der Kolk et al,[14] Roberts et al,[16] and Morbée et al[19] all reported good geometric accuracy for sCT in the lumbar or cervical spine, particularly for in-plane measurements.

Morbée et al[19] assessed geometric measurements on sCT (BoneMRI) of the lumbar spine in 30 participants against conventional CT. They found that measurements of pedicle width, spinal canal width, vertebral body height, width and length, and spinous process length on sCT were statistically equivalent to those on CT within a predefined equivalency margin of 0.5 mm. However, the neuroforamen length and maximal disc height showed larger deviations, potentially due to the differing patient positioning between MRI and CT scans.

Staartjes et al[13] demonstrated, in their proof-of-concept study of three cases, that pedicle screw trajectories and screw thickness were estimable based on lumbar sCT (BoneMRI) findings. They reported a mean absolute difference of 0.26 ± 0.24 mm for measurements such as vertebral body height and spinal canal diameter between sCT and conventional CT in one case.

van der Kolk et al[14] performed a detailed qualitative and quantitative assessment of BoneMRI-generated sCT of the cervical spine in 25 participants. They found good to excellent geometrical agreement with conventional CT for various distance measurements.

Roberts et al,[16] using a 3D CycleGAN for lumbar sCT from sagittal T1W MRI in 20 patients, found that measurements performed in the sagittal plane (e.g., vertebral body and disc heights) were generally accurate, with most showing less than 5% relative error compared with true CTs. However, pedicle width measurements reconstructed in the axial plane were considerably less accurate (relative error up to 34%), a limitation attributed to the nonvolumetric nature and poor out-of-plane resolution of the input sagittal MRI sequences.

 

MissingLinkText

Integration of sCT with Surgical Navigation and Robotics

The generation of high-fidelity sCT from preoperative MRI opens avenues for MRI-only workflows in navigated and robotic spinal surgery. Staartjes et al[13] explicitly stated that their sCT approach could facilitate surgical planning and intraoperative neuronavigation, potentially leading to “radiationless navigated surgery” and enabling the use of computer assistance based on preoperative imaging without the need for a separate planning CT. They also proposed the future possibility of radiationless robotic pedicle screw insertion using sCT. This aligns with the general trajectory of utilizing detailed preoperative imaging to enhance precision in robotic procedures and navigation potential.

 

MissingLinkText

 

MissingLinkText

Discussion

The generation of sCT from MRI using DL techniques represents a significant stride in medical imaging for spinal surgery. The primary motivations for this shift—reducing radiation exposure, streamlining workflows, and potentially lowering costs—are compelling. The literature increasingly supports the feasibility of producing sCTs with good image quality and geometric accuracy for various spinal applications.[11] [13] [14] [15] [16] [17] [18] [19] [20] The “usefulness” of AI in spine surgery, specifically through sCT, is evident in its potential to provide comprehensive diagnostic information (soft tissue from MRI, bone from sCT) in a single imaging session. This is valuable for preoperative planning, as demonstrated by studies showing accurate geometric measurements for the lumbar and cervical regions and for the direct diagnosis of conditions such as lumbar disc herniation and sacroiliitis.[11] [18] Customized patient care through sCT may enable more personalized surgical approaches tailored to individual patient anatomy and pathology. Furthermore, the prospect of integrating sCT with neuronavigation and robotic surgery could enhance intraoperative precision and safety without the need for preoperative planning CT or high-dose intraoperative CT.[4] [10] [13] Robust integration of surgical robotics and navigation systems remains a key area for development, promising improved guidance and precision.

Nevertheless, discrepancies still existed between the sCT and true CT.[11] [13] [14] [15] [16] [17] [18] [19] [20] Roberts et al highlighted that sCTs generated from 2D sagittal MRI slices can have poor accuracy for measurements in reconstructed axial planes (e.g., pedicle widths) owing to the poor out-of-plane resolution of the source MRI. van der Kolk et al[14] observed that certain complex pathologies like extensive ankylosis with vacuum phenomena or intraspinal bone appositions were not always accurately reproduced by sCT, and intraosseous pneumatocysts could be misinterpreted. The tendency of some sCT methods to reduce granular details in the trabecular bone was also noted. Resilience against MRI artifacts and the hazards of overfitting or inherent biases within DL models remain areas of continued research.

However, different DL architectures (U-Net, GANs, CycleGANs), input MRI sequences (T1W, T2W, specialized sequences such as T1W-MGE), and training strategies (paired vs. unpaired, 2D vs. pseudo-3D vs. 3D) yield varying results. For instance, CycleGANs allow training with more readily available unpaired data, but a direct comparison of their quantitative accuracy with methods trained on paired data with specific sequences (such as BoneMRI) requires careful consideration of the context. Fritz[20] raised the important question of whether sCT techniques truly add new diagnostic information beyond what might be interpretable from the source MR images themselves, or if they primarily offer an alternative, more familiar visualization format for bone. He suggests that DL-augmented approaches using multiple qualitative and quantitative MRI-based tissue parameters (e.g., from Dixon or multi-echo gradient-echo sequences) may be key to creating quantitatively accurate sCTs.

Despite encouraging developments, the widespread clinical application of DL-based sCT needs more robust and transferable results. While some studies, such as Bajger et al,[15] used large (800 patients) unpaired datasets, many validation studies are based on smaller, single-institution cohorts. The performance of DL models can be sensitive to variations in MRI acquisition parameters, scanner vendors, and patient populations that are not represented in the training data. Roberts et al[16] also noted the uncertainty of GANs in accurately representing pathology when trained on unpaired data. Assimilation into prevailing clinical practices and obtaining regulatory endorsements (e.g., FDA, CE mark) for diagnostic or surgical planning use are considerable hurdles that require validation and the demonstration of safety and efficacy.

Moreover, several challenges are notable for DL-based sCT clinical application. The need for expansive, diverse, and high-quality training datasets is paramount. Equity in healthcare access must be considered, as advanced technologies could potentially widen disparities if not implemented thoughtfully. Furthermore, as with all AI applications in healthcare, data privacy is a critical concern and must adhere to stringent regulations, such as HIPAA in the United States and GDPR in Europe.

Future studies should focus on multicenter studies and validation in more diverse patient groups, including those with significant pathologies or prior instrumentation. Future directions include further validation for a wider range of spinal pathologies and postoperative scenarios, improvement of through-plane resolution from 2D inputs or wider adoption of volumetric MRI sequences, and continued refinement of DL models to enhance accuracy and reduce susceptibility to artifacts. Research continues to enhance the efficiency and robustness of DL algorithms. The employment of techniques such as transfer learning and domain adaptation is gaining traction to overcome the constraints posed by limited datasets in specific institutions or for rare pathologies. The ultimate goal is to develop sCT tools that are not only accurate and reliable but also seamlessly integrate into clinical workflows to provide tangible benefits for both clinicians and patients.

 

MissingLinkText

Conclusion

DL-driven sCT from MRI constitutes a monumental leap in spinal surgery imaging. Current literature, including prospective studies and quantitative analyses, considers that this innovation holds substantial promise for streamlining preoperative planning, enhancing diagnostic capabilities for specific conditions, elevating intraoperative precision through better planning data for navigation and robotics, and, crucially, curtailing patient radiation exposure. Although challenges related to generalizability, artifact handling, and comprehensive validation across diverse pathologies and patient populations remain, the trajectory of development is strongly positive.

 

MissingLinkText

 

MissingLinkText

Conflict of Interest

None declared.

• References

• 1 Abumoussa A, Gopalakrishnan V, Succop B. et al. Machine learning for automated and real-time two-dimensional to three-dimensional registration of the spine using a single radiograph. Neurosurg Focus 2023; 54 (06) E16
• 2 Bahrami A, Karimian A, Arabi H. Comparison of different deep learning architectures for synthetic CT generation from MR images. Phys Med 2021; 90: 99-107
• 3 Cole KL, Findlay MC, Kundu M, Johansen C, Rawanduzy C, Lucke-Wold B. The role of advanced imaging in neurosurgical diagnosis. J Mod Med Imaging 2023;
• 4 Ahern DP, Gibbons D, Schroeder GD, Vaccaro AR, Butler JS. Image-guidance, robotics, and the future of spine surgery. Clin Spine Surg 2020; 33 (05) 179-184
• 5 Boulanger M, Nunes JC, Chourak H. et al. Deep learning methods to generate synthetic CT from MRI in radiotherapy: a literature review. Phys Med 2021; 89: 265-281
• 6 Chen G, Xu Z. Usage of intelligent medical aided diagnosis system under the deep convolutional neural network in lumbar disc herniation. Appl Soft Comput 2021; 111
• 7 Li W, Kazemifar S, Bai T. et al. Synthesizing CT images from MR images with deep learning: model generalization for different datasets through transfer learning. Biomed Phys Eng Express 2021; 7 (02)
• 8 Gotoh M, Nakaura T, Funama Y. et al. Virtual magnetic resonance lumbar spine images generated from computed tomography images using conditional generative adversarial networks. Radiography (Lond) 2022; 28 (02) 447-453
• 9 Hong KT, Cho Y, Kang CH. et al. Lumbar spine computed tomography to magnetic resonance imaging synthesis using generative adversarial network: Visual Turing Test. Diagnostics (Basel) 2022; 12 (02) 530
• 10 Burström G, Buerger C, Hoppenbrouwers J. et al. Machine learning for automated 3-dimensional segmentation of the spine and suggested placement of pedicle screws based on intraoperative cone-beam computer tomography. J Neurosurg Spine 2019; 31 (01) 147-154
• 11 Cao G, Li Y, Wu S. et al. Clinical feasibility of MRI-based synthetic CT imaging in the diagnosis of lumbar disc herniation: a comparative study. Acta Radiol 2024; 65 (01) 41-48
• 12 van Stralen M, van der Kolk B, Zijlstra F. et al. BoneMRI of the cervical spine: Deep learning-based radiodensity contrast generation for selective visualization of osseous structures. ISMRM 27th Annual Meeting and Exhibition; May 11–16, 2019; Montreal, QC, Canada; 2019
• 13 Staartjes VE, Seevinck PR, Vandertop WP, van Stralen M, Schröder ML. Magnetic resonance imaging-based synthetic computed tomography of the lumbar spine for surgical planning: a clinical proof-of-concept. Neurosurg Focus 2021; 50 (01) E13
• 14 van der Kolk BBYM, Slotman DJJ, Nijholt IM. et al. Bone visualization of the cervical spine with deep learning-based synthetic CT compared to conventional CT: a single-center noninferiority study on image quality. Eur J Radiol 2022; 154: 110414
• 15 Bajger M, To M-S, Lee G. et al. Lumbar spine CT synthesis from MR images using CycleGAN—a preliminary study. In: 2021 Digital Image Computing: Techniques and Applications (DICTA). Australia: IEEE Publication; 2021: 1-8
• 16 Roberts M, Hinton G, Wells AJ. et al. Imaging evaluation of a proposed 3D generative model for MRI to CT translation in the lumbar spine. Spine J 2023; 23 (11) 1602-1612
• 17 Oulbacha R, Kadoury S. MRI to CT synthesis of the lumbar spine from a pseudo-3D Cycle GAN. IEEE 2020; 17: 1784-1787
• 18 Jans LBO, Chen M, Elewaut D. et al. MRI-based synthetic CT in the detection of structural lesions in patients with suspected sacroiliitis: comparison with MRI. Radiology 2021; 298 (02) 343-349
• 19 Morbée L, Chen M, Herregods N, Pullens P, Jans LBO. MRI-based synthetic CT of the lumbar spine: geometric measurements for surgery planning in comparison with CT. Eur J Radiol 2021; 144: 109999
• 20 Fritz J. Automated and radiation-free generation of synthetic CT from MRI data: does AI help to cross the finish line?. Radiology 2021; 298 (02) 350-352

Publication History

Article published online:
03 September 2025