Investigating the Effects of Image Scaling Techniques in Radiographic Measurements of Spinal Alignment and Motion: A Comparative Analysis

Abstract

Radiographic measurements from spinal radiographs are crucial in many diagnostic and therapeutic decisions. However, widely used manual line drawing techniques in DICOM viewers are associated with significant errors, partly due to the unreliability of DICOM scale factors. A recently developed algorithm has been engineered to scale radiographs using assumed vertebral endplate widths (EPWs). Use of EPW to scale radiographs can eliminate the need to determine image magnification. This study was designed to (1) quantify the measurement error associated with the standard of care DICOM scale factor and (2) evaluate the efficacy of the EPW scaling algorithm. Previously collected cervical and lumbar lateral radiographs with a calibration marker of known size were used to collect radiographic measures of spinal alignment and spinal motion. Three different scale factors were used to acquire mm-based measurements: (1) the ground truth scale factor; (2) the DICOM scale factor; and (3) the EPW scale factor. DICOM and EPW scaled measurements were compared with respect to ground truth measurements. The DICOM scaled radiographic measurements demonstrated significantly more error compared to the EPW scaled radiographic measurements across multiple X-ray types: 19.6% vs. 9.7% in cervical neutral lateral (p < 0.001), 19.8% vs. 9.9% in cervical flexion/extension (p < 0.001), 35.0% vs. 7.5% in lumbar neutral lateral (p < 0.001), and 35.7% vs. 7.7% in lumbar flexion/extension (p < 0.001). The EPW algorithm provides adequate scaling accuracy for many clinical applications and is associated with considerably lower error than traditional DICOM scaled measurements. Clinicians should be aware of the potential scale factor inaccuracies in their imaging and be cautious when making diagnostic and therapeutic decisions based on improperly scaled radiographs.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-025-01680-7.

Introduction

Studies have investigated numerous biomechanical markers for the spine that are economically obtained from radiographs [1–4]. Such markers may be used to indicate the state of degeneration, deformity, or trauma, as well as for planning and monitoring various nonoperative and surgical treatments. Many of these measurements have shown direct correlations with patient outcomes, and payers use specific thresholds for certain radiographic measurements (alongside other criteria) to facilitate prior authorization for procedures [5]. Radiographic measurements are often referenced by clinicians in indications for surgery. For example, anterior–posterior translation of one vertebra with respect to another, measured in mm between lateral flexion and extension radiographs, is commonly used to provide evidence for the absence/presence of spinal instability, thereby providing objective reference points used in clinical guidelines [6, 7]. Static spinal alignment parameters (e.g., sagittal vertical axis) are also commonly reported in real-world units and are influential in surgical planning and assessments of spinal degeneration and deformity.

Manual line drawing with digital calipers in Digital Imaging and Communications in Medicine (DICOM) viewers has been a historical standard. It is widely used for obtaining radiographic measurements in many clinical workflows. By-hand measurements have been shown to be associated with significant measurement error (> 2°) and poor interobserver reliability for linear measurements (ICC = 0.151 for AP translation, ICC = 0.693 for rotation) [8, 9]. The strength of ICC agreement can be interpreted using the guidelines by Cicchetti et al.: poor, ICC < 0.40; fair, ICC ≥ 0.40 and < 0.60; good, ICC ≥ 0.60 and < 0.75; and excellent, ICC ≥ 0.75 [10]. Linear measurements may be further compromised by inaccurate DICOM scale factors used to convert pixels to millimeters (mm). DICOM scale factors are geometrically calculated and based on the distance between the X-ray source (S) and detector (D); it follows that inaccurate S-D machine calibration is one potential source of error given that even minor calibration inaccuracies can produce significant and potentially clinically consequential measurement errors relative to the size of the object being measured, especially as the field of view increases in size. Furthermore, patient positioning and body habitus alter the patient-to-detector distance and change the magnification of the X-ray, providing another potential source of DICOM measurement inaccuracy. Both calibration challenges and patient factors compromise the accuracy of geometrically calculated linear measurements made based on the DICOM scale factor. An inaccurate DICOM scale may only be remedied by limiting results to unitless percentages or by correcting the scale factor by using a calibrating marker of known size within the image.

Use of a calibration marker is the only routinely available method to produce clinical measurements from standard radiographs that reliably correspond to real-world metrics in raw numerical terms (percentages from the internal reference method mentioned above may be thought of as a unitless internal calibration without a predetermined denominator, where the reference object is scaled to 100). When used, a calibration marker should be placed at the same source-to-object distance as the anatomy of interest and in a consistent location across studies for comparison purposes. This marker can be used post-acquisition to specify the pixel-to-real-world unit (e.g., mm) scale factor [11]. Unfortunately, radiography technologists who do not routinely practice in an orthopedic setting may not be familiar with the use of a calibration marker or consistent in its use, even when aware. In high-volume settings, these markers are also sometimes viewed as disruptive to workflow and are used selectively, if at all. Calibration use is therefore highest in high-stakes protocol-driven acquisition settings such as clinical trials and/or surgical navigation, even though accurate measurements in real units are often desirable in a broader range of settings.

To address the demands of technologists, care providers, and payers for accurate real-unit measurements without disrupting workflow, an algorithm was developed that utilizes anterior–posterior vertebral endplate width (EPW) reference data from a large dataset obtained from radiographs that included physical calibration markers in the field of view. While it is hypothesized that utilizing the EPW algorithm for scaling lateral spinal radiographs will improve measurement accuracy when compared against radiographs scaled with the DICOM scale factor, the purpose of this study is twofold: (1) estimate the accuracy of measurements obtained from the current standard of care DICOM scale and (2) evaluate the efficacy of the EPW scaling algorithm by directly comparing radiographic measurements obtained from this method against measurements obtained by a ground truth method using an external calibration marker of known dimensions.

Methods

Experimental Design

A pooled analysis of spine radiographs, previously acquired from adults ages 18 to 80 for research purposes, was performed and included lateral lumbar and cervical spine radiographs from 29 sites and 39 unique radiography systems: 117 cervical neutral lateral (CNL), 109 cervical flexion/extension (CFE) pairs, 196 lumbar neutral lateral (LNL), and 192 lumbar flexion/extension (LFE) pairs were included. The spinal positioning for a neutral lateral radiograph is upright, for an extension radiograph arched backwards, and for a flexion radiograph arched forwards. Examples of each image type are included in Appendix 1 . All radiographs had been acquired with a calibration marker of known dimensions positioned in the mid-sagittal plane near the target anatomy to mitigate magnification effects (Fig. 1). Pearl IRB (www.pearlirb.com) found the research to be exempt according to 45 CFR 46.104(d)(4) Secondary Research Uses of Data or Specimens on 01/05/2024. IRB ID:2024–0008.

Fig. 1.

Lumbar radiograph with a calibration marker along the midsagittal plane annotated by SpineCAMP

For each radiograph, a fully automated and FDA-cleared intervertebral motion and alignment measurement technology (SpineCAMP™, Medical Metrics, Inc., Houston, TX) was used to automatically obtain the following measurements: average disc height, spondylolisthesis, and C2-C7 sagittal vertical axis (SVA) [12] in neutral lateral images and intervertebral translation from flexion/extension imaging. SpineCAMP measurements scaled with the calibration ring have been shown to be equivalent [13] to a previously validated, semi-automated technology (QMA™, Medical Metrics, Inc., Houston, TX). QMA has been independently shown to have sub-degree and sub-millimeter accuracy [8] as well as increased interobserver agreement of radiographic measurements with respect to by-hand measurements and CT assessments [9, 14, 15].

Linear measurements were initially acquired by the automated technology in pixels. Each analyzed image used square pixels with identical scaling in both the x- and y-directions. This results in no changes in angular measurements as image scaling changes. Since angular measurements are not affected by image scaling, only dimensional measurements were analyzed.

For this study, three different scale factors, with units of mm/pixels, were applied to the pixel measurements, resulting in three different scaled measurements in units of mm: (1) the ground truth scale factor using external calibration, (2) the DICOM scale factor obtained from DICOM image and radiography device settings, and (3) the EPW scale factor estimated by the novel EPW algorithm. For the ground truth scaling method, a calibration ring of known size was used to establish a pixel-to-mm scale factor. For the DICOM scaling method, the DICOM tag registered in the radiography machine’s settings was obtained from the metadata present in the image to establish a pixel-to-mm scale. For the EPW scaling method, an algorithm utilizing a large reference dataset of vertebral endplate widths was used to establish a pixel-to-mm scale for each image. Each of these methods is discussed in detail in the following sub-sections.

Establishing Ground Truth Measurements

A torus ring with a 30-mm middle diameter was used for the calibration marker in this study (Fig. 2). Metal discs and spheres have been used to scale radiographs [16]. The edges of discs and spheres can “wash out” as the exposure increases. In contrast, the middle diameter of the metal ring can be identified over a wide range of exposures (Fig. 3). This is important as the object used to scale X-rays should be as close to the spine as possible and at the same distance to the image plane as the spine. This can be accomplished by placing the scaling object against the patient’s back. However, in that position, there is no tissue to the left and right of the object to attenuate X-rays, and at the exposure setting required to see the vertebrae, a metal object posterior to the spine can wash out of the X-ray. It is for this reason that a metal ring was used. Even when the inner and outer edges of the ring wash out, the middle diameter can be identified and used to calibrate the X-ray.

Fig. 2.

The inner diameter (ID), outer diameter (OD), and middle diameter (MD) of the ring (left side), and the middle diameter when the ring is not in the same plane as the X-ray (right side)

Fig. 3.

Radiographic images of a metal ring and a United States quarter at four, increasingly high (images 1 to 4), radiographic exposure settings. At the highest exposure, the apparent diameter of the quarter is less than it is at the lowest exposure, yet the middle diameter of the ring is the same at all exposure settings

The underlying artificial intelligence (AI) used in SpineCAMP has been trained and validated to detect this ring and use the ring to calculate a pixel-to-mm scale factor for the image (30-mm ring diameter/ring diameter in pixels). All linear measurements from each image are scaled to mm based on the scale factor established using the ring. The scale factor is most accurate when the ring is placed at the same distance from the source and detector as the object of interest. Utilizing the pixel measurements established from SpineCAMP for that image, the calibration marker scaled measurements in mm, used as the ground truth for future comparisons, can be calculated utilizing Eq. 1.

$$GroundTruthLinearMeasurement\left(m,m\right)=AnatomicDimension\left(p,i,x,e,l,s\right)*CalibrationRingScaleFactor\left(\frac{\mathit{mm}}{\mathit{pixels}}\right)$$ 1

Investigating the Effect of Calibration Ring Positioning on Image Scaling

Establishing the True Width of the Vertebral Endplate

The superior L4 endplate width in the midsagittal plane was measured in 3D from the thin-slice CT scan using 3D Slicer [17]. Two different users measured the anterior–posterior endplate length in 3D, and the average of these two users was used as the best estimate of truth for the sensitivity analysis (BET) (Table 1). To most accurately measure the endplate, multiple views were used to identify the most anterior and posterior points along the midsagittal plane of the endplate. The distance between these two points was measured using the coordinates of the points.

Table 1.

Data to establish the best estimate of truth for the L4 superior endplate width

Reader	L4 superior endplate width measurement (mm)
Reader 1	31.06
Reader 2	31.60
Average	31.33

Sensitivity Analysis of Calibration Ring Positioning

Since measurements scaled with the calibration ring were used as the ground truth in this study, a sensitivity analysis was performed to investigate the effect calibration ring positioning has on radiographic image scaling. This was done using digitally reconstructed radiographs (DRRs) created from a thin-slice CT scan obtained from the New Mexico Decedent Image Database (NMDID) [18] (Fig. 4). The DRRs were created using Plastimatch DRR [19]. The anterior–posterior width of the superior endplate of the L4 vertebral body, in the mid-sagittal plane, was used as the target dimensional measurement. The aim was to investigate how the placement of calibration rings affects radiographic measurements of vertebral endplate width. This requires knowing the “true” width of the vertebral endplate.

Fig. 4.

L4 superior endplate measurements in the midsagittal plane in 3D Slicer

Creating Simulated X-rays and Making Endplate Measurements with AI Engine

Once the BET for the L4 superior endplate width was established, the CT volume was prepared with the calibration ring in various positions according to a comprehensive simulated X-ray acquisition protocol. A calibration ring, with dimensions similar to those used in clinical radiographs, was virtually added to the CT exam using 3D Slicer and positioned at various locations. Multiple DRRs were created for each calibration ring position, representing a range of clinically possible radiographic projections. The default position positions the ring perfectly in line with the sagittal plane, posterior to the L4 body and spinous processes. All other ring positions are made relative to the default ring position.

Once the simulated X-rays were created, they were automatically analyzed by the AI in SpineCAMP, which had been trained to recognize the calibration ring and determine the pixel-to-mm scale factor based on the calibration ring’s middle diameter of 30 mm. The AI also places landmarks on the vertebral corners. Using these landmark coordinates and the pixel-to-mm scale factor, the L4 superior endplate width in mm was calculated for each DRR and subsequently compared to the BET findings. The data from this experiment were analyzed to estimate variability in the calibration ring scale factor caused by variations in ring placement.

Acquiring Scaled Linear Measurements from the DICOM Scale Factor

In the absence of a calibration marker, the most common method of image scaling relies on the DICOM scale factor. A DICOM scale factor was extracted from the value specified in the Pixel Spacing (0028,0030) tag. If this tag was not specified, the Imager Pixel Spacing (0018,1164) tag value was used for the DICOM scale. These DICOM tags are calculated by the radiography machine and are based on pre-calibrated source, detector, and estimated patient distances (Fig. 5). The tags specify the physical distance between pixels according to the DICOM Standard PS3.3 2024 d and are the default scaling in DICOM image viewers, unless another preferred calibration scheme is selected, which is usually entered by the end-user. Once this scale factor is established, the exact linear measurements in pixels can be converted to millimeters (mm) using Eq. 2.

$$DICOMScaledLinearMeasurement\left(m,m\right)=AnatomicDimension\left(p,i,x,e,l,s\right)*DICOMScaleFactor\left(\frac{\mathit{mm}}{\mathit{pixels}}\right)$$ 2

Fig. 5.

An image acquisition diagram illustrates the key components used in calculating the DICOM scale factor

Acquiring Scaled Linear Measurements from an Endplate Width Algorithm

As an alternative to DICOM scaling, an algorithm utilizing average endplate widths was developed. Reference endplate widths from a proprietary, fully deidentified dataset comprised of 3000 cervical spine X-rays and 870 lumbar spine X-rays were gathered from pooled clinical trials data. The average EPWs established by this method were consistently found to be within the ranges described in the existing literature [20–23]. Real patient EPWs are influenced by several biological factors, with height and lean body mass showing a moderate correlation with vertebral EPW [24]. The algorithm employs a prioritization method to select the optimal level for comparison, which was determined during the algorithm’s development. The prioritization method was driven by the need to specify one reliable dimension that could be used to scale standard of care X-rays. Standard of care field of view, beam position, and anatomic variability known to be present in the cranial cervical spine and caudal lumbar spine influenced the design of the algorithm. Objective validation for the order of the prioritization was not performed. Differences in prioritization on the effect of the experiment could be tested in future research. The cervical and lumbar spine EPW reference data are listed in order of priority in Tables 2 and 3. In an image of the cervical spine, for example, if C5 is visible and analyzed, the superior endplate in pixels is utilized with an assumed superior endplate width of 15.11 mm, based on pooled reference values, to produce a scale factor for the image (distance from reference value in mm/measured endplate width in pixels). If C5 is unavailable, C6 will be used, and so on, in the order of the preassigned priority in the table. Quality control modules are in place to ensure that vertebral landmarks are accurate and will not be reported at levels with elevated uncertainty. These quality control modules, related to automated vertebral body labeling and landmarking, are protected by trade secrets. Once the EPW scale factor from the endplate width algorithm is established, linear measurements in pixels are converted to millimeters (mm) using Eq. 3.

$$EPWScaledLinearMeasurement\left(m,m\right)=AnatomicDimension\left(p,i,x,e,l,s\right)*EPWScaleFactor\left(\frac{\mathit{mm}}{\mathit{pixels}}\right)$$ 3

Table 2.

Anterior–posterior endplate widths used to specify mm-to-pixel calibration factor in lateral cervical spine X-rays in order of priority

Vertebra	Mean (mm)	Std Dev	Median	Lower 1%	Upper 99%
C5	15.11	1.88	15.0	11.5	19.9
C6	16.05	1.84	15.9	12.4	20.4
C4	14.92	1.82	14.8	11.4	19.8
C3	14.82	1.85	14.7	11.2	19.4
C7	16.52	1.82	16.4	12.9	20.9
C2	12.02	1.38	11.9	9.3	15.5

Table 3.

Anterior–posterior endplate widths used to set mm-to-pixel calibration factor in lateral lumbar spine X-rays in order of priority

Vertebra	Mean (mm)	Std Dev	Median	Lower 1%	Upper 99%
L4	33.05	3.14	32.9	26.4	40.4
L5	32.81	3.05	32.7	26.2	40.5
L3	33.89	3.15	34.1	26.7	41
L1	31.74	3.49	31.6	24.9	40.2
S1	32.25	3.23	32.1	25.4	40.2

Comparing Linear Measurements

A calibration ring was visible in every radiograph described in the “Experimental Design” section. The linear measurements obtained after scaling each image using the calibration ring were considered the ground truth. Absolute (mm and %) errors with respect to the ground truth were calculated for measurements scaled with the DICOM scale factor and EPW algorithm scale factor. Microsoft Excel and Stata (Ver 15, StataCorp LLC, College Station, TX) were used for all data analyses. A paired t-test was used to assess statistically significant differences between measurement types (α = 0.05). The workflow for this study is overviewed in Fig. 6.

Fig. 6.

Process map to establish linear measurements using three different scaling methods

Results

The percent absolute errors for measurements derived from the DICOM scale factor and the EPW scale factor relative to ground truth measurements were reported for each image (Table 4). The absolute error (mm) for each linear measurement obtained via the two methods was also reported (Tables 5 and 6). Statistically significant differences between measurements scaled with the EPW algorithm scale factor and the DICOM scale factor, with respect to the ground truth, were found in cervical neutral lateral, cervical flexion–extension, lumbar neutral lateral, and lumbar flexion–extension imaging (p < 0.001 for each).

Table 4.

% absolute measurement error by image type

Scaling type	Mean	Std. error	Std. Dev	95% CI
CNL EPW Scale (%)	9.74	0.25	5.89	[9.24 10.24]
CNL DICOM Scale (%)	19.57	0.17	3.98	[19.23 19.91]
CFE EPW Scale (%)	9.91	0.26	5.89	[9.40 10.43]
CFE DICOM Scale (%)	19.80	0.18	4.04	[19.45 20.16]
LNL EPW Scale (%)	7.51	0.19	5.93	[7.13 7.89]
LNL DICOM Scale (%)	35.01	0.31	9.64	[34.39 35.62]
LFE EPW Scale (%)	7.69	0.20	6.02	[7.31 8.07]
LFE DICOM Scale (%)	35.69	0.32	9.81	[35.06 36.31]

Table 5.

Static measurement absolute error: 117 cervical subjects, 196 lumbar subjects

Scaling type	N	Mean	Std. error	Std. Dev	95% CI
Cervical Avg DH EPW Scale (mm)	533	0.25	0.01	0.16	[0.24 0.26]
Cervical Avg DH DICOM Scale (mm)	533	0.51	0.01	0.17	[0.50 0.53]
Cervical Spondy EPW Scale (mm)	533	0.05	0.002	0.05	[0.05 0.06]
Cervical Spondy DICOM Scale (mm)	533	0.10	0.004	0.10	[0.09 0.11]
C2C7 SVA EPW Scale (mm)	115	2.57	0.24	2.55	[2.10 3.04]
C2C7 SVA DICOM Scale (mm)	115	4.93	0.23	2.50	[4.47 5.39]
Lumbar Avg DH EPW Scale (mm)	941	0.23	0.01	0.20	[0.21 0.24]
Lumbar Avg DH DICOM Scale (mm)	941	1.03	0.01	0.37	[1.01 1.06]
Lumbar Spondy EPW Scale (mm)	941	0.05	0.002	0.05	[0.04 0.05]
Lumbar Spondy DICOM Scale (mm)	941	0.22	0.004	0.13	[0.21 0.23]

Table 6.

Motion measurement absolute error: 109 cervical subjects and 192 lumbar subjects

Scaling type	N	Mean	Std. error	Std. Dev	95% CI
Cervical translation EPW Scale (mm)	506	0.08	0.004	0.09	[0.07 0.09]
Cervical translation DICOM Scale (mm)	506	0.16	0.01	0.13	[0.15 0.17]
Lumbar translation EPW Scale (mm)	953	0.04	0.001	0.05	[0.03 0.04]
Lumbar translation DICOM Scale (mm)	953	0.16	0.004	0.12	[0.15 0.17]

Violin plots were used to visualize the distribution of signed and absolute errors associated with each scaling method across the spectrum of imaging (Figs. 7 and 8). While the magnitude of the absolute error has been reported in previous tables, visualizing the signed error allows for potential bias (overestimating or underestimating) to be appreciated for each scaling method.

Fig. 7.

% error in image scaling for DICOM scale and EPW scale with respect to calibration marker scale

Fig. 8.

% absolute error in image scaling for DICOM scale and EPW scale with respect to calibration marker scale

The results of the sensitivity analysis done regarding the effect of calibration ring positioning on image scaling showed the most significant effects when the ring was moved along the anteroposterior axis. This causes the ring to appear larger or smaller in the image, leading to inaccurate estimations of the ring’s length. The effects on image scaling of positioning the calibration ring in various positions are shown in Table 7.

Table 7.

Ring location, measured L4 superior endplate width, and absolute error for simulated X-rays

Ring translation relative to default	Measured L4 superior endplate width (mm)	Absolute error (% BET)a
Ring in line (default)	31.39	0.20
100 mm away from the source	36.67	17.05
100 mm closer to source	27.03	12.85
250 mm anterior*	N/A	N/A
100 mm superior	32.02	2.20
50 mm inferior	32.20	2.78
Rotated 45° AP	32.57	3.97
Rotated 90° AP*	N/A	N/A

^aBET for L4 superior endplate has been measured to be 31.33 mm

*Ring not identified in this position

Discussion

The results of this study suggest that more effective methods may exist for radiographic calibration than the current ubiquitous DICOM scale factor method. The radiographic calibration marker method used in this study as the “ground truth” is a gold standard approach employed in settings where very high accuracy is critical, such as clinical trials or funded basic science research. However, it has drawbacks, including workflow limitations and technical training requirements. While any approach to calibration that does not use a calibration marker is based on an assumption, an approach that incorporates real-world data about anatomic sizes rather than simple geometric assumptions alone may provide higher accuracy. Such an approach requires minimal, if any, alteration in technologist workflow and is therefore more likely to be widely adopted clinically, thereby enhancing clinical decision-making for a broader range of patients. The assumed endplate width method described in this paper is one such method.

When comparing the two assumption-based approaches against the ground truth approach, both DICOM and EPW methods overestimate the size of endplate widths. The DICOM scale factor error exhibits a higher average error than the EPW scale factor, as shown in Figs. 6 and 7 (p < 0.001). From a clinical standpoint, overestimating a linear measurement by significant amounts may lead to false positives, which could contribute to unnecessary treatment and healthcare expenses. Although imperfect, the EPW algorithm provides an error profile that is smaller in magnitude than the DICOM scale factor calibration method, potentially leading to more reliable radiographic measurements.

The assumed EPW algorithm was established based on the average vertebral size in a large proprietary dataset. The same automated analysis tool used in this study was employed to provide superior endplate measurements in millimeters from lateral radiographs, with a calibration of known size within the field of view. The images used to establish the EPW algorithm dataset were not utilized in the assessment of error related to scale factors in this study. The accuracy of these assumptions has not yet been systematically evaluated in real-world cases. As the actual size of the vertebral endplate diverges from the assumed endplate width, the error associated with the assumed endplate width algorithm will increase.

While the algorithm performs well in most cases, in 76 out of 608 (12.5%) linear measurements, there was an absolute error greater than 15% between the calibrated measurement and the assumed measurement scaled using the EPW algorithm. By comparison, the mean absolute error in all categories of study in which the DICOM method was used was greater than 19% in the cervical spine and greater than 35% in the lumbar spine. Measurement errors from any method may be clinically significant depending on the application. Clinicians should consider the effects of improper scaling when applying radiographic measurements, understanding the impact of measurement error on clinical decision-making. Utilizing a calibration marker will provide substantially more accurate linear measurements in situations where substantial measurement error is likely to lead to adverse outcomes.

The accuracy of linear measurements made in millimeters (mm) plays a crucial role in several clinical use cases, such as the diagnosis of spinal instability. The existing literature generally defines the diagnosis of spinal instability by the presence of 3–4 mm of intervertebral translation between flexion and extension [25, 26]. For lumbar spinal stenosis patients, the presence of instability is an indication for spinal fusion in addition to the neural decompression that is required to alleviate the stenosis. Given the significant differences in time, difficulty, complication potential, and expense between procedures (decompression alone vs. decompression plus fusion), the presence or absence of instability has become a focal point in clinical decision-making. False negatives can lead to the patient not getting the needed stabilization and potentially subsequent revision procedures. In contrast, false positives can lead to a more invasive procedure associated with an increased risk profile and higher healthcare expenses. Insurers’ prior authorization considerations often rely heavily on objective data to justify procedures, as it is easier to incorporate into approval algorithms and is perceived as a more substantive basis for approval or denial.

In the medicolegal realm, the 5th Edition of the AMA’s Guides to the Evaluation of Permanent Impairment utilizes mm-based linear measurements to determine the presence of the alteration of motion segment integrity (AOMSI) [27]. AOMSI plays a significant role in determining the impairment rating of an injured individual. This impairment rating, along with other factors, largely determines the settlement an individual may receive following an accident. Thus, the accuracy of the linear measurements in this scenario is crucial and must withstand the scrutiny of a legal setting.

Another clinical use case that requires accurate linear measurements is assessing and planning spinal implants and deformity correction. While CT often supports such planning, radiographs are used in rod planning for long spinal fusions. They are also used to assess device performance and track longitudinal changes [28]. Aside from the preoperative use case, postoperative monitoring of constructs and revisions may be influenced by inaccuracies in linear measurements. It is worth noting that increasingly popular biplanar systems incorporate spatial data and processing capabilities to substantially reduce or eliminate the scale factor concerns raised in this research. However, traditional single-plane digital X-rays with long cassettes (still commonly used for deformity assessment and treatment planning) remain subject to the problem discussed here.

The results of this study suggest the EPW scale factor is more reliable and associated with less error than the DICOM scale factor. The algorithm published here relies on the accurate specification of vertebral body landmarks, which can be reliably delivered automatically in a workflow via a clinically available software as a medical device (SaMD). The emergence of automated technology to assist with radiographical measurements has shown promise in driving improved accuracy and efficiency in clinics. A previous study assessing the performance of a machine learning and computer vision algorithm demonstrated statistically equivalent accuracies to manual measurements made by experienced surgeons, while taking only 0.14 s per radiograph on average to measure lumbar lordosis [29]. Semi-automated technology has been associated with increased agreement among different users. In an analysis of 30 sets of digital lumbar flexion–extension radiographs, interobserver reliability utilizing a semi-automated measurement tool was significantly higher than traditional by-hand methods (rotation 0.962 vs. 0.693, AP translation 0.862 vs. 0.151, change in anterior disc height 0.862 vs. 0.373, and change in posterior disc height 0.730 vs. 0.300) [9]. A fully automated analysis tool with proper linear measurement calibration could provide increased accuracy, interobserver reliability, and time savings.

While there has been limited research on the errors associated with utilizing the DICOM scale to make linear measurements, some studies have been performed that further support the findings in this paper. Rampersaud et al. (2006) assessed the accuracy of linear measurements in a cervical spine sawbone positioned 24 cm from the receiver (representative of the shoulder width of a 70-kg man) at four centers using four commercially available digital imaging viewing programs. For CT, measurements are shown to be valid (accurate within 1 mm of ground truth). For plain film X-rays, linear measurements were 5–8 mm (~ 40%) greater than the ground truth (p < 0.0001) [30]. Building upon these findings, Ravi et al. [31] demonstrated a significant positive correlation between magnification error in lateral lumbar X-rays and BMI (p < 0.0001). Thus, patients with high BMI (increasing the distance between the target anatomy and receiver) are at an elevated risk of receiving erroneous linear measurements from lateral spinal X-rays [31]. These studies support the findings of significant inaccuracies of DICOM scaled measurements in a large cohort of cervical and lumbar lateral X-rays.

Despite a thorough analysis of several radiographic scaling methods, this study has limitations. One such restriction surrounds the reference data used to develop the EPW algorithm. Due to the deidentified nature of the reference and test data, it cannot be guaranteed that this reference data, and thus the EPW algorithm, sufficiently represents the general population, despite the inclusion of a relatively large number of studies. Another limitation is that the images used in the analyses were only available as de-identified with unknown demographic data, aside from the age restrictions common to all commonly available de-identified datasets (e.g., no pediatric data was included). Nevertheless, vertebral sizes are relatively fixed by adulthood, and the algorithms incorporate error-checking mechanisms. While both the EPW reference data and imaging in the study come from pooled clinical trials data, which are often designed to represent a multiethnic population, the exact information regarding sex and race cannot be explicitly stated.

Another limitation in this study is that the analysis was limited to the cervical and lumbar spines. Since the automated AI measurement tool has only been cleared by the FDA for analysis of the cervical and lumbar spine, no EPW algorithm has been developed for the thoracic spine. The development of a thoracic spine EPW algorithm and a thorough assessment of linear measurements in the thoracic spine would be necessary to confirm that the findings in this study apply to the thoracic spine as well. Finally, while the EPW algorithm performs well in most cases, patients with very large or very small vertebral endplates may receive inaccurate approximations for their endplates, resulting in inaccurate linear measurements. For example, a subject with an L4 superior endplate of 26.4 mm (representing the lower 1% of the reference data) will overestimate linear measurements by 20.1% when utilizing the assumed EPW of 33.05 mm. As mentioned above, the algorithm still achieves better performance, on average, than the DICOM approach.

Conclusions

Reliable linear measurements from radiographs are essential for accurate clinical decisions. This study highlights significant scaling errors using the ubiquitously employed DICOM scale factor that average over 19% in cervical imaging and over 35% in lumbar imaging, with potentially harmful effects on clinical outcomes. A new automated approach, based on computer vision identification of endplate margins and estimation of endplate width using a large, calibrated reference dataset, reduces measurement error to less than 10% on average, resulting in an approximately 50–70% error reduction without a substantial change in workflow. When optimal accuracy is required, such as for implant sizing, preoperative planning, or custom rod length calculations, a calibration marker is crucial to avoid errors and ensure proper treatment. However, for general clinical use, the EPW algorithm may provide an enhancement to clinical care over measurements made only with DICOM calibration.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contribution

The conceptualization and design of this study were performed by Trevor F. Grieco, John A. Hipp, and Seth C. Coomer. The methodology was established by Seth C. Coomer, Trevor F. Grieco, and Alexander N. Merkle. Data collection and data analysis was performed by Seth C. Coomer. The first draft of the manuscript was written by Seth C. Coomer, and all authors contributed, commented, and edited versions of the manuscript. All authors read and approved the final manuscript.

Declarations

Ethics Approval

Pearl IRB (www.pearlirb.com) found the research to be exempt according to 45 CFR 46.104(d)(4) Secondary Research Uses of Data or Specimens on 01/05/2024 IRB ID: 2024–0008.

Consent to Participate

This study was deemed exempt according to 45 CFR 46.104(d)(4). All imaging and data in this study was fully deidentified prior to its availability to the authors.

Consent for Publication

This study was deemed exempt according to 45 CFR 46.104(d)(4). All imaging and data in this study was fully deidentified prior to its availability to the authors.

Competing interests

Seth C. Coomer, John A. Hipp, and Trevor F. Grieco are employed by Medical Metrics Inc. Alexander N. Merkle, Vikas V. Patel, and Pierce D. Nunley are paid consultants for Medical Metrics Inc. John A. Hipp owns stock in Medical Metrics Inc. Trevor F. Grieco owns stock in Medical Metrics Diagnostics. Vikas V. Patel has a consulting relationship with Cerapedics Inc., Zygofix, SpineWelding AG, Shukla, Mainstay Medical, Globus Medical Inc., Orthobond Corporation, and SI-BONE Inc. Vikas V. Patel has received past grant funding from Simplify Medical and Mainstay Medical. Pierce D. Nunley has a consulting relationship with SeeALL, Camber Spine, Accelus, Providence Medical, Neo Spine, Spineology, Centinel Spine, Intrinsic Therapeutics, Spinal Elements NG Medical, Regeltec, 3Spine, Redrock, SAIL Fusion, SST, and Nevro. Pierce D. Nunley has received grant funding from Globus Medical, Providence Medical, Centinel Spine, SAIL Fusion, and 3Spine. Pierce D. Nunley has board membership with SeeALL.