The potential role of synthetic computed tomography in spinal surgery: generation, applications, and implications for future clinical practice

Shreya Sankar; Jake Michael McDonnell; Stacey Darwish; Joseph Simon Butler

doi:10.31616/asj.2024.0197

Asian Spine J > Volume 18(6); 2024 > Article

Sankar, McDonnell, Darwish, and Butler: The potential role of synthetic computed tomography in spinal surgery: generation, applications, and implications for future clinical practice

Review Article

Asian Spine Journal 2024;18(6):913-922.

Online first: December 31, 2024

DOI: https://doi.org/10.31616/asj.2024.0197

The potential role of synthetic computed tomography in spinal surgery: generation, applications, and implications for future clinical practice

Shreya Sankar^1,², Jake Michael McDonnell^1,³

, Stacey Darwish^1,⁴, Joseph Simon Butler^1,⁴

¹National Spinal Injuries Unit, Mater Misericordiae University Hospital, Dublin, Ireland

²School of Medicine, Royal College of Surgeons, Dublin, Ireland

³Trinity Centre of Biomedical Engineering, Trinity College, Dublin, Ireland

⁴School of Medicine, University College Dublin, Dublin, Ireland

Corresponding author: Jake Michael McDonnell, National Spinal Injuries Unit, Department of Orthopaedics, Mater Misericordiae University Hospital, Dublin 7, Dublin, Ireland,
Tel: +353-852802958, E-mail: jakemcdonnell@rcsi.com

Received May 23, 2024 Revised September 5, 2024 Accepted November 1, 2024

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Computed tomography (CT) is widely used for the diagnosis and surgical treatment of spinal pathologies, particularly for pedicle screw placement. However, CT’s limitations, notably radiation exposure, necessitate the development of alternative imaging techniques. Synthetic CT (sCT), which generates CT-like images from existing magnetic resonance imaging (MRI) scans, offers a promising alternative to reduce radiation exposure. This study examines the emerging role of sCT in spinal surgery, focusing on usability, efficiency, and potential impact on surgical outcomes. This qualitative literature review evaluated various sCT generation methods, encompassing traditional atlas-based and bulk-density models, as well as advanced convolutional neural network (CNN) architectures, including U-net, V-net, and generative adversarial network models. The review assessed sCT accuracy and clinical feasibility across different medical disciplines, particularly oncology and surgery, with potential applications in orthopedic, neurosurgical, and spinal surgery. sCT has shown significant promise across various medical disciplines. CNN-based techniques enable rapid and accurate generation of sCT from MRI scans, rendering clinical use feasible. sCT has been used to identify pathologies and monitor disease progression, suggesting that MRI alone may suffice for diagnosis and planning in the future. In spinal surgery, sCTs are particularly useful in visualizing key anatomical features like vertebral dimensions and spinal canal diameter. However, challenges persist, especially in visualizing complex structures and larger spinal regions, like the lumbar spine. Additional limitations include inaccuracies stemming from surgical implants and image variability. The application of sCT technology in spinal surgery holds great promise, improving diagnostics, planning, and treatment outcomes. Although further research is required to improve its precision, it offers a viable alternative to traditional CT in many clinical contexts, with the potential for broader application as the technology matures.

Keywords: Computed tomography; Magnetic resonance imaging; Convoluted neural networks; Spine surgery

Introduction

Computed tomography (CT) is an imagingmodality that generates cross-sectional images of patients through strategic deployment and capture of rotating X-rays. Invented in 1967 by Sir Godfrey Hounsfield, the CT scan has revolutionized the medical field, becoming one of the most reliable and widely utilized imaging procedures in healthcare [1]. The first CT scan, conducted in 1971 in Wimbledon, England, accurately visualized a cystic astrocytoma, demonstrating the technology’s diagnostic potential [2]. Since then, CTs have consistently demonstrated clinical efficacy across various specialties and pathological states [3].

In spine surgery, CTs are extensively utilized in both diagnosis and treatment, with a study demonstrating the essential role of CTs in identifying midspinal fractures [4]. Furthermore, CT technology has enhanced surgical precision, driving better outcomes in various spine surgery subspecialties, including minimally invasive, pediatric, and general adult procedures [5,6]. CT is also used extensively intraoperatively, with the development of the O-arm (Medtronic, Minneapolis, MN, USA) CT allowing surgeons to make informed, real-time decisions in the operating theater based on the CT images without compromising patient stability. Notably, the O-arm was found to be exceptionally useful for intraoperative navigation, achieving a 99% accuracy rate for pedicle screw placement [7]. However, CT has limitations in imaging certain patient populations, such as those with metal implants [8]. Intraoperative use is also limited by complexity, protocol constraints, and patient positioning issues [9]. Moreover, heavy reliance on CT for postoperative imaging has exacerbated radiation exposure, with patients averaging two scans in the 5 years following surgery [10]. The risks associated with excessive radiation exposure, coupled with the complexities and intraoperative limitations of conventional CT, underscore the need for an alternative imaging technique that addresses these shortcomings while maintaining image accuracy and reliability. A promising solution is synthetic CT (sCT) imaging.

sCT is a form of tomography generated from MRI scans, eliminating the need for the time, logistics, and radiation required for traditional CT scans. An important consideration is sCT’s capability to reduce intraoperative radiation exposure through precise surgical planning. Studies have shown that sCT can accurately depict bony degenerative changes and pathologies, informing critical decisions on screw thickness and pedicle placement during surgery [11]. Although intraoperative applications have been explored, the current consensus holds that sCT cannot be used independently due to several limitations, including the requirement for precise patient positioning, field depth issues during scanning, and restricted applicability in complex cases [9,12]. These limitations are particularly relevant in patients with metal implants, where the accuracy of both CT and MRI are already compromised. Consequently, generating precise sCTs from these flawed original scans is challenging. Therefore, further research is required to further develop the sCT technology and enable seamless intraoperative adoption. Furthermore, given the relative novelty of sCT, there is a paucity of research on its long-term effects and broader implications for medical and surgical practice. The purpose of this review was to comprehensively explore the efficacy, potential applications, and ramifications of sCT technology in the management of spinal injuries.

Synthetic Computed Tomography

Traditional methods

Traditionally, sCT generation has relied on two primary methods: bulk-density and atlas-based models. The bulk-density approach involves segmenting MRI scans into air, soft-tissue, and bone compartments, each assigned a corresponding electron density, from which an sCT is then derived [13]. In contrast, atlas-based models employ a target scanning technique, where a reference MRI (atlas MRI) is paired with a corresponding CT (atlas CT) and then registered to a patient-specific MRI, enabling the necessary adaptations to generate an sCT [13]. Additionally, patch-based sCT generation offers another approach, wherein small sections of the patient’s MRI are compared to corresponding sections of atlas MRIs. The most similar matching sections are then selected and used to construct the sCT image [13]. Other studies have explored the usage of dictionary learning-based sCT construction, a method similar to atlas-based models, wherein a patient’s MRI is compared to a standard MRI, analyzing their differences, and identifying two standard CTs with similar differences. These CTs are then paired and adapted using electron densities derived from the MRIs to create an sCT image. Forest-based methods have also been explored, utilizing several “decision trees” to find the most efficient way to calculate electron densities from paired MRIs and CTs. The sCTs are then generated by aggregating the results from all decision trees [14]. Although these methods have been relatively successful, their potential for inaccuracy, lengthy processing times, and high computational complexity limit their practical implementation in many clinical settings [13].

Deep-learning methods

Convolutional neural networks

Recently, machine learning techniques have emerged as a promising approach for sCT generation, particularly convolutional neural networks (CNN) and generative adversarial networks (GAN). CNNs operate by aggregating a dataset of images/inputs, and identifying and analyzing specific features through a process called convolution [15]. This involves applying a kernel filter to scan sections and generating feature maps that highlight areas where particular features are present. The resulting feature maps enable the CNN to produce an output image summarizing the detected features [16]. Consequently, CNNs can integrate multiple CT images and scans to create a cohesive sCT. By inputting numerous patient CTs and images into a CNN, an sCT can be derived (illustrated in Fig. 1). Researchers demonstrated the efficacy of CNN-generated sCTs in a study involving 11 patients with head and neck cancers, where high-quality sCTs were produced from a single regular CT and a cone-based CT scan [17]. In a separate study, investigators used a fully convolutional network (FCN), a modified CNN architecture, to construct sCTs from patient MRIs. By segmenting the MRI into three two-dimensional (2D) slices and incorporating calculations from brain tissue and cerebral matter, the FCN-generated sCTs exhibited enhanced resolution. The study concluded that this approach holds promise for clinical applications, warranting further investigations [18].

A more complex V-net architecture, building upon existing CNN technology, has also shown promise. The architecture features two key components: a compression pathway and a mirroring decompression process. The compression pathway processes input information at various resolutions, learns spatial transformations and applies them to generate a new synthetic image. While creating sCTs, the compression pathway analyzes CT sections, converts each into corresponding sCT segments, and concatenates these segments from each layer into a single, cohesive sCT image [19].

Generative adversarial networks

To improve the accuracy of sCT scans, some studies have explored the addition of an adversarial loss term to traditional CNNs, which serves to compare sCTs to real CTs and minimize the differences [14]. These GANs consist of a generator and discriminator, which generate data and distinguish between genuine and ingenuine data [19]. This allows for more accurate sCT generation, as 2D slices from a paired MRI–CT scan is manipulated and converted into a three-dimensional model, while the discriminator simultaneously accounts for the mistakes. The process can be further augmented using dense cycle GANs, wherein a denser generator records more nuanced variations and “multiscale information” between MRI–CT pairings [14]. These technologies have expanded the viability of using sCT in healthcare settings and provide a basis for further research on developing novel software to address the more complex needs of sCT generation.

Early Clinical Application and Results in Various Medical Fields

Given the novelty of sCT technology, existing research is limited, primarily involving small patient cohorts. Therefore, the literature falls short of conclusively establishing the efficacy and broad applicability of sCT in clinical contexts. Further research is required to substantiate its relevance and effectiveness.

Oncology

The first United States Food and Drug Administration-approved commercial sCT technology was employed in a cohort of 25 patients with prostate cancer [20]. The software used in the study created sCTs with only approximately 1 mm distortion, and no more than 2 mm differences in cone-based CT generation compared to traditional CT scans. The scans were accurate enough for physicians to monitor other important metrics, such as rectal and bladder filling [20]. Another study employed a GAN-based CNN to create sCT images from cone-based CTs and evaluated their accuracy in 30 patients with pancreatic cancer. The image quality and dose calculation capabilities of the technology were evaluated using various metrics, including mean absolute error (MAE), spatial nonuniformity, signal-to-noise ratio, cross-correlation, and structural similarity indices comparing original planning CTs, cone-beam CTs, and the three types of generated sCTs (U-net, GAN, GAN with attention gates). Results showed that sCT using GAN and attention gates had the lowest values for each metric, with no significant differences compared to traditional CT [21]. Furthermore, sCT accurately displayed the sizes and positions of various organs [21].

Radiotherapy

Although sCT has proven effective in terms of overall image clarity, its application in patients undergoing very minute or highly detail-oriented surgeries remains understudied. One study investigated sCT accuracy using CNN-based MRI-to-sCT conversion technology in patients who had previously undergone skull resectional surgeries. The researchers calculated MAE and dice similarity coefficient, assessing bone overlap between original CTs and generated sCTs [22]. Following strict registrations of sCTs to CTs, only minor differences were observed in dose-based calculations, attributed to mechanical variations. The study concluded that sCTs were adequately accurate for use as a reference for future neural radiotherapy.

sCT has also demonstrated utility in abdominal radiotherapy, as evidenced by a study converting abdominal MRIs of five patients to sCTs. The MRIs were first classified into fat, high-density tissue, spine, air, and lung segments. Each of these sections was then converted to sCT using a “fuzzy-c network” that identified and delineated structures in the scans, followed by manual contouring to generate sCTs. The resulting sCTs had an average difference of 0.2% (with an outlier at 0.8%) compared to the original MRIs, indicating their potential for future treatment planning in the gastrointestinal tract [23].

The efficacy of existing sCT technology in oncology and radiotherapy is evident, with the generated scans yielding images comparable to conventional CT scans. However, there is a need to improve neuroradiological sCTs, particularly in delineating fine structural details.

Orthopedic Surgery Application and Efficacy

sCT images generated from MRI scans have also demonstrated high accuracy in orthopedic studies, as evidenced by a paper examining its application in hip morphology. MRI and CT scans of 30 patients with prostate cancer were obtained, and U-net software was used to create sCTs. The mean error and surface distance between corresponding locations on sCTs and the original CTs were then compared. Two senior radiologists and one radiologist-in-training identified various anatomical landmarks on the images. The results showed that the sCTs accurately depicted the femur and pelvis with a root-mean-square error of only 0.8 mm, and the bone geometry was deemed accurately reconstructed [24]. However, the accuracy of sCT was compromised in patients with certain hip pathologies, as degenerative changes had reduced image clarity in the original CT/MRI scans. Furthermore, segmenting the scans into 3 mm sections for comparison resulted in imprecise identification of various features [24]. Although using thinner sections would increase processing time due to the greater number of sections to be analyzed, it is hypothesized to improve image accuracy.

MRI-based sCT generation was further studied in 30 patients with suspected inflammatory sacroiliitis. Participants underwent MRI and CT scans on the same day, and U-net technology was applied to the MRIs. The resulting sCTs were segmented into 1 mm-thick slices and compared to the original CTs for accuracy. Additionally, the sCTs were compared to TI-weighted MRIs to quantify precision differences. Two radiologists also evaluated the sCTs and original images to identify structural lesions in both sacroiliac joints [25]. The accuracy, specificity, and sensitivity of sCTs for detecting erosion, sclerosis, and ankylosis were all at least 91% (except for erosion sensitivity, which was 78%), with p-values ranging from <0.001 to 0.49. Notably, the percentages for sCT were either higher or comparable to those obtained with T1-weighted MRIs for all the studied disorders. Moreover, radiologists accurately identified pathologies in sCTs at the same level as in the original CTs, demonstrating the clinical adaptability of sCTs. However, the researchers acknowledged study limitations, including the small sample size and potential selection bias.

The orthopedic applications of sCT were further reinforced by a study that converted lower arm MRIs to sCTs using GAN and U-net software. The process involved matching real CT scans and MRIs, followed by MRI manipulation to generate sCTs, and subsequent calculation of MAE, dice similarity, and surface-to-surface distance between the original and sCTs [26]. The mean surface distance for sCT to CT comparison was 0.53 mm, while its inverse was 0.43 mm, indicating an overall satisfactory correspondence. However, errors were observed around boundaries, attributed to suboptimal MRI–CT registration [26] (Fig. 2). Distinguishing between bones and tendons also proved difficult, and the generation of the wrist and its associated structures was not entirely accurate, potentially attributed to signal distortion from the MRI and CT scans due to the wrist’s distal location. Consequently, the study concluded that sCTs hold promise for radial and ulnar orthopedic applications. However, additional refinement is needed to improve the visualization of the wrist and other smaller structures in the arm.

This comprehensive review demonstrates the overall efficacy of sCT in orthopedic applications. However, the existing studies also revealed major shortcomings, such as compromised sCT generation and clarity due to thick CT segments, small sample sizes, and inadequate precision in visualizing complex anatomical structures.

Neurosurgery Application and Efficacy

sCT also holds promise in brain surgery, particularly in Gamma Knife (GK) radiosurgery. Given that CT-derived dose algorithms outperform MRI-derive ones, the need for easier generation of CTs, or sCTs, is amplified in this field. Researchers investigated sCT usability using a U-net architecture to create synthetic images from 30 sets of paired MRI/CT scans. Twenty-four sets were used for machine learning training, while six sets were reserved for evaluating output precision using MAE, mean error, and mean squared error metrics. Results showed that the generated sCTs achieved high accuracy (95.4%–98.6%) for most soft tissues in the head region, except at tissue-air interfaces [27]. The authors attributed the minor lapses in sCT precision to suboptimal MRI–CT matching and registration, the presence of metal implants, anatomical differences resulting from previous surgery, and inherent MRI distortions. Despite these errors, the study concluded that their impact was small enough, and overall, the results validate the credibility and feasibility of using sCTs in GK radiosurgery planning.

A similar study employing GAN architecture to create sCTs for neurosurgical guidance corroborated these findings. The study analyzed images from 75 patients, categorized into four groups: moving MRI, moving CT, stationary MRI, and stationary CT. Each MRI–CT pairing was synthesized to produce both sCTs and synthetic MRIs (sMRIs), which were evaluated using metrics such as dose errors and dice coefficients. Although the generated sCTs accurately replicated structures, CSF, and brain parenchyma found in original CTs, notable precision lapses were observed [28]. The sCTs displayed high levels of uncertainty in extracranial soft-tissue, and structures were portrayed in poorer resolution, affecting the readability [28]. While the brain was reliably reproduced, much of the extraneural contents had high error values, indicating that sCTs are not yet ready for stand-alone use in neurosurgery. The issues are largely attributable to sizable MRI slice thickness, which restricts detailed image analysis, and uncertainties in MRI–CT registration [28]. Additionally, the scans analyzed in this study concerned deep brain pathologies, which likely contributed to the heightened accuracy of neural structures compared to extracranial contents. Notwithstanding these errors, sCT images were found suitable for emulating brain deformation.

Traditional atlas-based methods of sCT generation have also proven effective for neural imaging. In a study, scans of ten patients with brain cancer were used to generate sCTs, with an additional patient’s MRI serving as the atlas image. The patient MRIs and CTs were registered and run through the atlas software to create sCTs [29]. The synthetic images produced from MRIs were highly accurate, specifically in visual image analysis of the ventricles and frontal lobe. Further, density measurements for CSF, gray matter, and white matter on the sCTs, expressed in Hounsfield units (HU), closely matched those of traditional CTs, corroborating the reliability of this approach. While this research focused on sCT generation to improve MRI–CT registration, the findings imply that sCTs are accurate enough to be used for future neural and neurosurgical studies.

Preliminary Evidence for Synthetic Computed Tomography Implementation in Spine Surgery

sCT has also demonstrated benefits in spinal surgery. A study evaluated the accuracy of sCTs through image analysis and surgical robot integration, utilizing a female cadaver. Traditional CT, MRI, and sCT scans (generated using BoneMRI deep-learning technology) were obtained and analyzed. The results showed remarkable precision, with a mean surface distance of −0.02±0.05 mm and a mean absolute surface distance of 0.24±0.05 HU between the original CT and sCT [11]. These scans then acted as a guide for a surgical robot to insert screws in the cadaver.

The robot utilized the traditional sCT for left-sided pedicle screw placement, and the original CT for right-sided placement. The tip distances (1.74±1.1 mm and 2.36±1.6 mm, p=0.24), tail distances (1.93±0.88 mm and 2.81±1.03 mm, p=0.07), and angular deviations (3.2 ± 2.05 mm and 4.04±2.71 mm, p=0.53) of sCT and CT were similar, underlining the potential application of sCT technology for surgical guidance and navigation. Another study employed a patch-based machine learning method for generating sCTs. The study included eight patients (to train the algorithm), one prospective elective patient, and two healthy volunteers [30]. The spinal canal diameter and anterior and posterior vertebral body heights were measured for vertebrae L3–L5, with a mean absolute difference of 0.26±0.24 mm reported. The sCT also identified spondylolisthesis in one of the volunteers, allowed for accurate screw insertion planning by the researchers, and correctly measured lordosis and spinal diameter. These findings indicate the potential for diagnostic and therapeutic application of sCT in clinical settings.

Two additional studies have corroborated the high accuracy of sCTs in the lumbar spine. One study utilized publicly available 18 MRI–CT pairings, segmenting each MRI scan to isolate vertebral bodies and pedicles, and then converting them using a GAN architecture. Accuracy testing involved identifying six structures on each vertebra. Additionally, U-net technology was applied to the separated vertebral bodies and pedicles. The results showed an average MAE of 125.65±10.07 HU, an average dice error of 0.83±0.16, and a landmark error of 2.2±1.4 mm [31]. Although the low dice error and landmark error indicate the potential for spine surgery applications, the MAE was relatively high, likely due to the overall larger structures in the lumbar spine. The positive association between structure size and MAE may become a significant limitation of sCT applications if the high error continues to impact reliability. Another GAN-based study on 20 MRI–CT pairs assessed the accuracy of sCT in lumbar spine measurements, with two neurosurgeons and two radiologists evaluating four parameters: intervertebral disc height, mid-vertebral body height, lumbar lordosis, and right and left pedicle width. The relative errors were between 1% and −7% (the outliers were L5/S1, at −10%), 1% and 3% (the outlier was S1, at 8%), 10%, and between −19% and 34%, respectively [12]. The values for the last metric corroborate the previous study’s findings on larger structures having higher errors, reinforcing concerns about sCT’s reliability in relation to anatomical size. Except for pedicle width, the results demonstrate the applicability and reliability of sCT in lumbar spine measurements. The high correlation between sCT measurements and original CT structural measurements, combined with the interrater agreement between neurosurgeons and radiologists, supports the utility of sCT. However, the study acknowledges limitations, including decreased precision due to degenerative changes, positioning variability [12], and errors in the generation algorithm, some of which went undetected during machine learning, particularly affecting pedicle width. However, future research is expected to mitigate these shortcomings.

Promising results have also been observed in the cervical spine, where researchers employed an “echo sequence,” a method similar to patch-by-patch registration, to create sCTs from MRIs. Using MRI–CT pairings from 25 patients, the study created sCTs that were then evaluated by a multidisciplinary team. A neurosurgeon, musculoskeletal radiologist, and neuroradiologist identified several anatomical structures in the synthetic scans, while an additional musculoskeletal radiologist conducted a visual comparison between sCTs and the original CTs to detect differences. Qualitatively, the sCTs exceeded expectations by a large margin, creating images on par with the original CTs [32]. The generated sCT scans demonstrated improved resolution (specifically of the lower spine), wider views, clearer blood vessel disfigurements, and better image quality around metal implants compared to the original CTs. The clinicians were also able to reliably identify most anatomical structures and landmarks. These findings were supported by the quantitative metrics: an MAE of 80.05±6.12 HU, dice coefficient of 0.84±0.04, and structural similarity index of 0.86±0.02, indicating “good to excellent agreement” with CT scans [11]. The researchers attributed any errors to positional variation and minor, addressable issues.

Notably, sCT has shown great potential in spine surgeries with relevance to the ossification of posterior longitudinal ligament (OPLL). A study generated sCTs from MRIs of 22 patients, and two radiologists were asked to assess OPLL extent in the original MRI, the sCT, and a separately captured CT scan. Notably, sCT demonstrated higher sensitivity for detecting OPLL extent compared to MRI, with values of 90% versus 47% and 93% versus 63% for the two radiologists, respectively. This improvement stemmed from the combined input from MRI and CT, providing clearer visualization of OPLL. Additionally, there was a positive correlation between structures in the CT and sCT [33]. These results were further corroborated by a study employing CNN-based architecture to diagnose OPLL from MRI scans, based on features identified by radiologists. The technology correctly selected all scans with OPLL. The overall accuracy was 98%, with a sensitivity of 85% and specificity of 98% [34]. These two studies illustrate the clinical capabilities of sCT in imaging OPLL.

The available evidence suggests the potential of sCT technology to facilitate diagnosis, particularly in OPLL-related surgery, structure identification, and image analysis owing to its superior quality (particularly for robot-guided surgeries and in the cervical spine). However, sCT currently has a notably high MAE for larger spinal structures, further exacerbated by changes in device position, computational mistakes, and in some cases, pre-existing patient pathologies.

Future Outlooks and Limitations

The preliminary findings from the aforementioned studies indicate a promising potential for sCT in medical applications. sCT has shown reliability in various fields, including oncology, radiology, neural and spinal surgery, and orthopedics. Current technologies used to generate sCTs are accessible and easily utilizable by medical professionals, ranging from simpler atlas-based and patch-by-patch methods to advanced artificial intelligence-driven software. As the field evolves, sCT may offer significant benefits, including reduced costs and streamlined workflows for healthcare facilities and patients. Furthermore, integrating sCT with emerging deep-learning technologies, augmented reality, and visualization techniques can help improve healthcare efficiency.

Despite its efficiency, sCT faces several limitations that need to be addressed [11,12,24,26–28,31] (Table 1). First, the accuracy of sCT heavily relies on precise MRI–CT image registration. This process is susceptible to several factors, such as patient and instrument positioning, capturing of exact locations/measurements, and human error which can occur at any point during the image processing. Mitigating these errors requires highly precise MRI–CT pairing technology, which poses significant development and implementation challenges. The impact of the original MRI’s Tesla score on the sCT’s accuracy remains understudied. Current literature does not appear to report a clear trend between the usage of 1.5 T or 3 T MRIs and sCT sensitivity. Both field strengths have produced relatively accurate sCTs. For example, 3 T MRIs have enabled precise recreation of spinal structures and accurate diagnosis of pathologies, specifically OPLL [11,33,34]. Similarly, 1.5 T MRIs have generated near-perfect sCTs, facilitating reliable identification of structures and pathologies by surgeons [32]. These 1.5 T-based sCTs also demonstrated low MAEs and high structural correlation [12,30]. Notably, both 1.5 T and 3 T MRIs appear equally susceptible to limitations such as positional variation and algorithmic errors, leading to decreased precision, particularly for overly large or very complex structures. Although 1.5 T and 3 T MRIs demonstrate similar performance in sCT generation, further research is warranted to investigate optimal MRI protocols for various spine pathologies. Specifically, exploring the relationship between MRI intensity and pathology visualization could ensure consistent sCT accuracy. Further, several studies have highlighted challenges in producing high-resolution images of smaller, complex bony structures, distal features, and those affected by degenerative pathologies, compromising the readability and reliability of sCTs. This limitation may become a significant barrier to sCT implementation in the future, particularly given the rising prevalence of degenerative spinal diseases [35]. Unless advanced software capable of producing high-resolution images of degenerated anatomy is developed, sCT’s applicability may be limited in its scope. Studies have also revealed that larger structures in the spine tend to exhibit higher MAEs on the sCT scans. This finding suggests that sCT may be more effective when applied only to the cervical spine or when size limits are imposed on the structures being imaged. Further research comparing MAEs between the cervical spine and lumbar spine would help elucidate the relationship between structure size and sCT accuracy. Nevertheless, the development of advanced software and error-reducing technologies is anticipated to mitigate these inaccuracies.

One of the major hypothesized benefits of sCT lies in its intraoperative application in spinal surgery, where it could alleviate the burden of traditional O-arm CT machines. Specifically, sCT promises to reduce radiation exposure while providing the necessary information to surgeons. However, sCT technology has not yet advanced to this caliber of utilization beyond cadaveric studies, as it has shown considerable issues in accurately reconstructing images with degenerative changes and containing larger structures, both of which are common scenarios in intraoperative CT usage. Consequently, implementing sCT in complex procedures like adult deformity surgeries and lumbar fusions may involve trade-offs: reduced radiation, time, and costs, but potentially compromised image accuracy. Another critical consideration is the challenge of comparing generated sCTs to original CTs for precision when the source MRIs were acquired on a different day. Active degeneration can alter the position and appearance of structures, complicating direct comparisons. Despite this, sCTs have demonstrated reasonable image quality, allowing surgeons to identify structures with acceptable accuracy. However, more robust studies are required before the clinical pilot application of sCT technology in spine surgery.

This review is greatly constrained by the scarcity of existing research on sCT, stemming from its novelty and restricted availability for academic purposes. Current literature is characterized by small patient cohorts and notable demographic exclusions, such as pediatric patients and individuals with previous surgeries or metal implants, due to anticipated accuracy issues. These exclusions limit the generalizability of available sCT evidence. Furthermore, the comparison between sCTs and contrast-enhanced CT remains unexplored, hindering a comprehensive understanding of sCT’s potential applications in spinal procedures. Future studies should evaluate sCT’s usability more specifically in spinal surgery and pathologies.

Conclusions

sCT is an emerging technology enabling the generation of CT scans from existing MRI scans. This paper reviewed the various forms of generation, clinical applications, and implications of sCT imaging, and found cautiously optimistic support for its potential benefits in spinal surgery. While, current literature demonstrates sCT’s relative effectiveness in visualizing structures, diagnosing various pathologies, and approximating original CTs, its limitations include imprecision in depicting larger, distal, and complex structures, and reduced accuracy due to changes in machine positioning, MRI slice thickness, and algorithmic errors. Therefore, sCT is not yet an entirely reliable alternative to traditional CTs. Further investigation of these constraints in pre-clinical models is needed before clinical adoption. Nevertheless, the realm of radiation-free intraoperative navigation in spinal surgery may prove within grasp in forthcoming years.

Key Points

Pre- and perioperative spine surgery is associated with significant exposure to computed tomography (CT)-radiation.
Synthetic CT can be constructed from existing MRI images and has the potential to reduce radiation exposure for patients.
However, current limitations hinder its widespread adoption into clinical practice.

Notes

Conflict of Interest

No potential conflict of interest relevant to this article was reported.

Author Contributions

Conceptualisation: JMD, SD, JSB. Data curation: SS. Formal analysis: SS, JMD. Funding acquisition: NA. Methodology: JMD. Project administration: JMD, SD, JSB. Visualisation: JMD, SD, JSB. Writing–original draft: SS, JMD. Writing–review & editing: JMD, SD, JSB. Final approval: all authors.

Fig. 1

Graphical flowchart depicting the process of synthetic computed tomography (CT) generation using machine learning through analysis of multiple magnetic resonance imaging (MRI) and CT.

Fig. 2

(A, B) Image depicting differences between the original computed tomography (CT) and reconstructed synthetic CT (sCT) by color—coding the distance between a point on the original CT and the corresponding closest point on the sCT. Adapted from Zijlstra F, et al. arXiv [Preprint] 2019 Jan 24. https://doi.org/10.48550/arXiv.1901.08449 [26].

Table 1

The limitations of sCT as reported in current literature, categorized by its medical, structural, and logistical shortcomings

Medical	Structural	Logistical
Degenerative disease affected clarity around the degenerated structures [12,24].	Distal and complex structures were less accurate [26].	Improper registration of the MRIs and CTs used to develop sCT technology [27,28].
Metal implants were displayed in poorer image quality [27].	Larger structures had higher MAEs [12,31].	Positional variability between the CT and MRI [11,12]
		Thicker MRI slices used for input into sCT software may cause less precise generation [28].
		Surgical or degenerative changes between the time of CT and MRI reduced comparability of sCT to the original CT [27].

sCT, synthetic computed tomography; CT, computed tomography; MRI, magnetic resonance imaging; MAE, mean absolute error.

References

1. Schulz RA, Stein JA, Pelc NJ. How CT happened: the early development of medical computed tomography. J Med Imaging (Bellingham) 2021;8:052110.

2. International Society for Computed Tomography. Half a century in CT: how computed tomography has evolved [Internet] [place unknown]: International Society for Computed Tomography. 2016 [cited 2024 Jan 10]. Available from: https://www.isct.org

3. Rubin GD. Computed tomography: revolutionizing the practice of medicine for 40 years. Radiology 2014;273(2 Suppl): S45–74.

4. Rajasekaran S, Vaccaro AR, Kanna RM, et al. The value of CT and MRI in the classification and surgical decision-making among spine surgeons in thoracolumbar spinal injuries. Eur Spine J 2017;26:1463–9.

5. Adamczak SE, Bova FJ, Hoh DJ. Intraoperative 3D computed tomography: spine surgery. Neurosurg Clin N Am 2017;28:585–94.

6. Otomo N, Funao H, Yamanouchi K, Isogai N, Ishii K. Computed tomography-based navigation system in current spine surgery: a narrative review. Medicina (Kaunas) 2022;58:241.

7. Mattei TA, Fassett DR. Letter to the editor: the O-arm revolution in spine surgery. J Neurosurg Spine 2013;19:644–7.

8. Ghodasara N, Yi PH, Clark K, Fishman EK, Farshad M, Fritz J. Postoperative spinal CT: what the radiologist needs to know. Radiographics 2019;39:1840–61.

9. Holly LT, Foley KT. Intraoperative spinal navigation. Spine (Phila Pa 1976) 2003;28(15 Suppl): S54–61.

10. Patel VV, Andersson GB, Garfin SR, Resnick DL, Block JE. Utilization of CT scanning associated with complex spine surgery. BMC Musculoskelet Disord 2017;18:52.

11. Davidar AD, Judy BF, Hersh AM, et al. Robot-assisted screw fixation in a cadaver utilizing magnetic resonance imaging-based synthetic computed tomography: toward radiation-free spine surgery. Illustrative case. J Neurosurg Case Lessons 2023;6:CASE23120.

12. 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;23:1602–12.

13. 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;89:265–81.

14. Lei Y, Harms J, Wang T, et al. MRI-only based synthetic CT generation using dense cycle consistent generative adversarial networks. Med Phys 2019;46:3565–81.

15. Currie G, Hawk KE, Rohren E, Vial A, Klein R. Machine Learning and deep learning in medical imaging: intelligent imaging. J Med Imaging Radiat Sci 2019;50:477–87.

16. Sarigul M, Ozyildirim BM, Avci M. Differential convolutional neural network. Neural Netw 2019;116:279–87.

17. Chen L, Liang X, Shen C, Jiang S, Wang J. Synthetic CT generation from CBCT images via deep learning. Med Phys 2020;47:1115–25.

18. Li G, Bai L, Zhu C, Wu E, Ma R. A novel method of synthetic CT generation from MR images based on convolutional neural networks. In: Proceedings of the 11th International Congress on Image and Signal Processing: BioMedical Engineering and Informatics (CISP-BMEI); 2018 Oct 13–15; Beijing, China. Piscataway (NJ): IEEE. 2019 https://doi.org/10.1109/CISP-BMEI.2018.8633142

19. 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.

20. Tyagi N, Fontenla S, Zhang J, et al. Dosimetric and workflow evaluation of first commercial synthetic CT software for clinical use in pelvis. Phys Med Biol 2017;62:2961–75.

21. Liu Y, Lei Y, Wang T, et al. CBCT-based synthetic CT generation using deep-attention cycleGAN for pancreatic adaptive radiotherapy. Med Phys 2020;47:2472–83.

22. Lerner M, Medin J, Jamtheim Gustafsson C, Alkner S, Siversson C, Olsson LE. Clinical validation of a commercially available deep learning software for synthetic CT generation for brain. Radiat Oncol 2021;16:66.

23. Hsu SH, DuPre P, Peng Q, Tome WA. A technique to generate synthetic CT from MRI for abdominal radiotherapy. J Appl Clin Med Phys 2020;21:136–43.

24. Florkow MC, Willemsen K, Zijlstra F, et al. MRI-based synthetic CT shows equivalence to conventional CT for the morphological assessment of the hip joint. J Orthop Res 2022;40:954–64.

25. Jans LB, 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;298:343–9.

26. Zijlstra F, Willemsen K, Florkow MC, et al. CT synthesis from MR images for orthopedic applications in the lower arm using a conditional generative adversarial network. arXiv [Preprint] 2019 Jan 24 https://doi.org/10.48550/arXiv.1901.08449

27. Yuan J, Fredman E, Jin JY, et al. Monte Carlo dose calculation using MRI based synthetic CT generated by fully convolutional neural network for gamma knife radiosurgery. Technol Cancer Res Treat 2021;20:15330338211046433.

28. Han R, Jones CK, Lee J, et al. Deformable MR-CT image registration using an unsupervised, dual-channel network for neurosurgical guidance. Med Image Anal 2022;75:102292.

29. Roy S, Carass A, Jog A, Prince JL, Lee J. MR to CT registration of brains using image synthesis. In: Proc SPIE Int Soc Opt Eng; 2014 Mar 21; [Epub]. https://doi.org/10.1117/12.2043954

30. Staartjes VE, Seevinck PR, Vandertop WP, van Stralen M, Schroder ML. Magnetic resonance imaging-based synthetic computed tomography of the lumbar spine for surgical planning: a clinical proof-of-concept. Neurosurg Focus 2021;50:E13.

31. Oulbacha R, Kadoury S. MRI to CT synthesis of the lumbar spine from a pseudo-3D cycle GAN. In: Proceedings of the 17th International Symposium on Biomedical Imaging; 2020 Apr 3–7; Iowa City, USA. Piscataway (NJ): IEEE. 2020 pp 1784. –7. https://doi.org/10.1109/ISBI45749.2020.9098421

32. van der Kolk BB, Slotman DJ, 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.

33. Jeong HS, Park C, Kim KS, Kim JH, Jeon CH. Clinical feasibility of MR-generated synthetic CT images of the cervical spine: diagnostic performance for detection of OPLL and comparison of CT number. Medicine (Baltimore) 2021;100:e25800.

34. Shemesh S, Kimchi G, Yaniv G, Harel R. MRI-based detection of cervical ossification of the posterior longitudinal ligament using a novel automated machine learning diagnostic tool. Neurosurg Focus 2023;54:E11.

35. Martin BI, Mirza SK, Spina N, Spiker WR, Lawrence B, Brodke DS. Trends in lumbar fusion procedure rates and associated hospital costs for degenerative spinal diseases in the United States, 2004 to 2015. Spine (Phila Pa 1976) 2019;44:369–76.