Computed Tomography of the Spine

The number of computed tomography (CT) examinations performed has been on the rise for decades [1‐3]. Increases in clinical application are related to technical developments, wider availability, and physician and patient demands [1, 2]. CT is at the forefront of imaging for multiple purposes, spanning from regular oncologic staging to acute imaging in the emergency trauma setting, contributing significantly to accurate diagnosis, optimized patient management, and improved treatment; however, the use of CT is inherently accompanied by exposure to ionizing radiation, which may cause radiation-induced malignancies [4, 5]. More specifically, it is assumed that about 2% of future cancer cases will be attributable to current application of imaging techniques [3, 6]. Thus, a general principle is to keep radiation exposure as low as reasonably achievable (ALARA principle) [7, 8]; however, in daily clinical routine, CT-related radiation exposure still varies considerably within and across institutions, given that well-defined and ubiquitous reference standards are frequently missing [9, 10]. One relevant aspect is that general recommendations are hard to determine considering the various scanner models and technologies, which may exert an impact on radiation exposure during scanning of different body regions.

CT examinations of the spine are performed for different indications including fracture detection and trauma evaluation, assessment of degenerative changes, postoperative complications, and guidance of interventional procedures, such as periradicular infiltration (PRI) [11‐13]. Particularly in musculoskeletal and neuroradiology departments, spinal CT constitutes a large proportion of the daily workload. Evidently, the most effective way to reduce CT-related radiation exposure is to use the technique only when the clinical value outweighs the risks and costs. Aside from that, various developments have emerged on both the acquisition and the reconstruction sides to achieve an optimized trade-off between image quality (IQ) and radiation exposure [14, 15]. Among others, dose reduction techniques include shielding of radiosensitive organs [16], beam-shaping filters [17] and, most importantly, the optimization of acquisition parameters, including tube voltage (kV), tube current, voxel size and slice thickness. Tube current is expressed either directly (as mA) or indirectly in terms of tube current-time product (as mAs). Different parameter combinations can lead to entirely different IQ at the same radiation dose.

In clinical CT examinations, radiation exposure is normally controlled via tube current modulation, which nowadays is usually achieved by means of automatic exposure control (AEC) [18, 19]. By adjusting the tube current to the patient’s habitus in the axial plane and along the z‑axis, a considerable dose reduction can be achieved. Tube current reduction results in a decreased patient dose, as the amount of generated X‑ray photons is directly proportional to the tube current [20]; however, image noise increases exponentially, mostly driven by Poisson noise. Exponential noise increase can be avoided by pulse-width modulation of tube current or X‑ray flux. This technique, termed sparse-sampling CT, reduces projections generated during a 360° gantry rotation while the dose for each individual projection remains constant. Technical implementations are challenging as they require high voltage fast switching electrical elements or fast shuttering of the X‑ray source [21‐23].

Any dose reduction technique usually comes at the cost of increased image noise and artifacts. Adequate image reconstruction techniques can mitigate these drawbacks and are therefore a major component of CT dose reduction.

Filtered back projection (FBP) is an analytical reconstruction algorithm relying on the exact mathematical relation between measured projection and reconstructed image data. The speed and robustness of FBP have made it the workhorse of CT reconstruction for decades [15, 24]; however, the assumption of noise-free data and the amplification of noise by the filter severely limit the quality of FBP-reconstructed CT images. In contrast, iterative reconstruction (IR) techniques can reduce image noise through iterative filtering or close to reality physical modeling of the data acquisition process [15]. IR algorithms can be categorized into three stages. Image domain-based reconstruction was the first clinically approved technique in 2009 and features high reconstruction speed; however, noise reduction is limited due to the rather simple iterative denoising only in image space. Hybrid IR algorithms, such as iDose (Philips Healthcare, Best, The Netherlands), ASIR (GE Healthcare, Milwaukee, WI, USA), or SAFIRE (Siemens Healthineers, Erlangen, Germany), feature increased noise reduction and reconstruction time through iterative filtering of both projection and image data. The last stage is represented by model-based IR algorithms (MBIR), which use advanced models in an iterative process of backward and forward projections. They achieve the highest level of noise reduction and ensuing dose reduction but are also computationally most demanding.

The emergence of artificial intelligence (AI) bears great potential for further dose reduction at almost all stages of CT imaging. On the acquisition side, AI-based algorithms have been developed to automatically position the patient at the isocenter using an infrared camera, select the scan range for the required anatomical coverage, or determine tube parameters, in order to optimize patient exposure [25]. On the reconstruction side, AI can be applied to denoise reconstructed images or even perform the reconstruction itself [15, 25]. One common approach is to train a convolutional neural network (CNN) with (simulated) low dose (LD) or artificially noise-enhanced data to reconstruct standard dose (SD) high-quality CT images [26‐29]. The application of a CNN is computationally inexpensive once it is trained and validated, when compared to MBIR algorithms. Besides improved IQ, noise reduction, and artifact reduction, reconstruction speed is thus another benefit of AI-based reconstruction. The use of AI in CT imaging will further reduce the required dose; however, study results usually cannot be readily generalized as AI networks are trained on specific datasets. This lack of generalizability must first be addressed to translate AI-based dose reduction to the patient and the ultimate clinical set-up.

Dose reduction techniques are increasingly being used for spinal CT; however, they have not been systematically reviewed yet. Therefore, we aimed to systematically review dose reduction and clinical applications of LD-CT on the spine. Our objective was to determine the degree of clinically achievable dose reduction and the effects on IQ, diagnostic confidence, and patient outcomes. For this purpose, we focused on four major clinical indications: (i) vertebral fractures (VF) and spinal trauma, (ii) spinal degeneration, (iii) perioperative evaluation and (iv) interventional procedures.

In this article, dose reduction techniques for spinal CT were systematically reviewed. We included 40 studies representing the most common clinical indications. Comparison of LD and SD was most frequently performed between modified tube settings and reconstruction techniques.

For evaluation of VF and spinal trauma, achieved dose reductions ranged from 6–71%. The majority of studies reduced the dose by at least 50% while maintaining overall diagnostic performance and confidence. Besides tube settings and reconstruction techniques, patient positioning and decreasing overlapping scan regions were approaches to reduce exposure. For evaluation of degenerative spine disease, dose reductions without a negative effect on diagnostic performance and acceptable IQ range of 35–50%. Although not consistently investigated across all included studies, overall dose reduction potential tended to be higher for more advanced reconstruction techniques and nonobese patients. Highest dose reductions were achieved for perioperative evaluation and interventions. The reported values for perioperative evaluation ranged from 75% to 95% without negatively affecting the clinical value of the images. For interventional procedures, dose reductions ranged from 34% to 93%, largely depending on the dose reduction approach as well as type and targeted phase of the procedure. The majority of those studies even achieved a dose reduction of > 70% while maintaining sufficient IQ for planning and guidance.

Dose reduction in spinal CT has in a large part been achieved by modifying tube settings while ensuring acceptable IQ using advanced reconstruction techniques. While these advances in LD-CT have been effectively enabled by new software, current and future developments in CT hardware will very likely increase dose reduction. Sparse-sampled CT enabled an additional dose reduction by a factor of 2 or more in simulation studies of the spine [71‐74] as well as in other indications [87]. Clinical translation can be expected once the required X‑ray tube technology is available for patient examinations [23]. Up to now, spinal applications of spectral CT have mainly been restricted to artifact reduction [88]. The clinical introduction of photon counting CT (PCCT) can be considered a new era for CT imaging, also with respect to radiation exposure [89, 90]. This innovative technology is expected to further improve IQ mainly due to reduction of electrical noise and artifacts, thus enabling dose reductions. Furthermore, it will potentially advance quantitative capabilities of spinal CT, such as more accurate BMD measurements and bone marrow quantification via material decomposition. Another emerging technique on the brink of clinical translation is AI which can be expected to bring additional dose reductions to spinal CT affecting both acquisition (e.g., via optimized patient positioning or scan volume selection) and reconstruction (e.g., via CNNs trained on low-quality LD and high-quality SD data) [25‐29].

In conclusion, considerable dose reduction in spinal CT can be realized by general approaches, such as tube setting modifications and advanced image reconstruction, but can be further increased through specific techniques for certain applications. Additional dose reduction up to 50% with comparable image quality can be expected from the clinical transition of novel acquisition and reconstruction techniques in the upcoming years.