Abstract
Background Artificial intelligence has previously demonstrated the capability to interpret cervical spine imaging. The present study aims to identify whether deep learning can be harnessed to triage patients into operative and nonoperative groups based on imaging findings, an intricate process complicated by the high rates of background findings.
Methods A deep-learning algorithm was trained to segment the spinal canal and spinal cord on 100 axial cervical spine magnetic resonance images to generate a biomarker for cervical stenosis, defined as the minimum difference in cross-sectional area between these structures. After training, the model and its biomarker were tested against a cohort of 147 consecutive patients evaluated in the outpatient setting by a group of board-certified neurosurgeons at our institution for complaints related to their cervical spines.
Results The mean minimum difference in cross-sectional area between the spinal canal and spinal cord for our cohort was 35.90 ± 25.00 mm2 for patients who ultimately underwent surgical decompression and 48.55 ± 33.52 mm2 for patients who did not (P = 0.005). Using this biomarker, the model distinguished between surgical and nonsurgical patients with relatively high accuracy (area under the curve 0.79).
Conclusion The present work describes a proof-of-concept model for triaging patients with cervical stenosis into surgical and nonsurgical cohorts based on imaging findings. In doing this, we aim to provide an additional supportive metric for referring providers to consult when considering whether to refer patients to spine surgery.
Clinical Relevance Neck pain is a common presenting symptom in the primary care setting, yet determining which patients could benefit from a surgical consultation is often a nuanced task. Deep learning has the potential to assist with the radiographic component of the referral process by identifying markers suggestive of operative candidacy, providing an objective metric for referring providers to consider when deciding whether surgical consultation is warranted. This may help triage patients by expediting referral of operative patients to spine surgeons and redirecting nonoperative patients to physiatry, physical therapy, or other conservative management.
Level of Evidence 3.
- cervical stenosis
- deep learning
- cervical spine MRI
- surgical candidacy
- imaging-derived biomarker
- spinal cord compression
Introduction
Cervical spine evaluations comprise a substantial number of primary care consultations each year.1,2 While current evidence affords these patients a variety of options, including exercise programs, medical therapies, diagnostic studies, and procedural interventions, the first question often posed to their referring physicians is whether the patient’s complaints warrant referral to a surgeon.
In many instances, a patient’s clinical history and physical examination findings will prompt a physician to obtain magnetic resonance images (MRIs) of the cervical spine. Given the cost of obtaining an MRI, considerable efforts have been made in the primary care setting to ensure they are ordered only for the appropriate clinical indications.3,4 Nevertheless, a significant number of MRIs are ordered each year, and the referring provider is often tasked with determining whether the results can be managed alone or require the input of a surgical subspecialist. However, owing to the relative nuance of interpreting spine MRIs, referring providers often rely on radiological reports to determine the degree of pathological severity in a given image. In some cases, a referral to a surgeon may be suggested by the reading radiologist in the text of the radiology report. In others, the ordering provider must make that determination using the qualifiers within the report, a few of which speak to whether or not a particular finding is treatable with surgery.
The ability of artificial intelligence to accurately interpret and autonomously segment the spine is well established in the literature.5–8 The present study builds on this foundation by evaluating whether these imaging-derived data can be leveraged to support clinical decision-making in referral and triage contexts. In the current study, we present a proof-of-concept deep learning model that uses imaging-derived features from axial, noncontrast T2-weighted cervical spine MRIs to triage patients into surgical and nonsurgical cohorts using a single quantitative biomarker to provide an additional objective metric that referring providers can consult when assessing a patient’s need for surgical consultation. We then discuss the key components of the model and propose that this tool, and others like it, could enhance decision-making and optimize the referral process to spine surgeons by offering quantifiable data to supplement clinical judgment.
Methods
IRB and Ethical Statement
The present study was conducted under the World Declaration of Helsinki: ethical principles for medical research involving human subjects.9 The present study was conducted according to the rules and regulations of our institution, as approved by the Institutional Review Board (IRB #16-000196). Waiver of informed consent was granted for the purposes of data collection by our institution’s IRB.
Manual Segmentation for Training Network
Our machine learning model was trained on manually segmented MRIs. One hundred T2 axial MRIs scans were randomly selected from our institution’s picture archiving and communications system. Images were downloaded as DICOM files, anonymized by removing header information, converted to NIFTI format, resampled to 256 × 256 pixels in the plane of acquisition, and, finally, histogram matched to a common template to standardize the maximum and minimum pixel values to 1 and 0, respectively. Following this preprocessing, clinically trained human expert raters delineated cervical canals and spinal cords in the axial plane on these scans using ITK-SNAP, an open-source application for medical image segmentation, as shown in Figure 1.5 Segmentations were stored as NIFTI files alongside the preprocessed training data for deep learning training.
Segmentations of the cervical spinal cord (red) and cervical spinal column (green) generated automatically using our machine-learning model on magnetic resonance images on a nonsurgical patient (A and B) and a surgical patient (C and D).
Deep Learning Model Training
Our ensemble contained 3 U-Nets trained on image patches extracted from preprocessed axial T2-MRIs. The first U-Net was trained using patches 160 × 160 px, the second using patches 128 × 128 px, and the third using patches 96 × 96 px. Patches were extracted in the plane of acquisition at regular strides of length 80 px, 64 px, and 48 px, respectively. Extracted image patches, along with the corresponding extracted region from the segmentation images, were augmented 3 times using affine transformations and flipped up-down as well as left-right. The collection of augmented and original patches and segmentations was then used to train the network.
The 3 networks in our ensemble were identical in all hyperparameters except the input image size. The relevant hyperparameters were a kernel size of 3 × 3 px with a stride of 1 px in each dimension, Rectified Linear Unit (ReLU) activation, and Glorot initialization for all convolutional layers. The starting channel depth was 16 for all networks, and our networks had 5 downsampling and upsampling units. Each unit consisted of a max pooling block using a 2 × 2 px kernel sandwiched between 2 convolutional blocks. Skip connections were added in accordance with the standard U-Net dogma. Channel depth was doubled with each downsampling and halved with each upsampling according to standard U-Net dogma as well. The final layer of all networks was a sigmoid to constrain the output between 0 and 1 for segmentation purposes. We also used dropout regularization with a probability of 0.25 and the Adam optimizer with a fixed learning rate of 1e-5 for training, alongside the Dice coefficient as a loss function. Segmentations generated using all 3 networks (96 × 96 px, 128 × 128 px, and 160 × 160 px) in the ensemble were combined using pixel-wise majority voting to generate the final segmentation. The results presented here are based on 2 ensembles: the first trained to segment spinal canals automatically and the second trained to segment spinal cords automatically in the axial plane on cervical spine MRI.
Biomarker Generation
Canal and cord areas were measured in square millimeters at all cervical levels except the most cranial 20% of each image due to the presence of brain regions that were irrelevant to our research question. The difference in area between the spinal canal and spinal cord was computed at every level, and the minimum value for this difference in area along the entire cervical spinal canal served as our biomarker. Segmentation failure for a particular axial slice was defined when irrelevant regions of the spinal canal (ie, the medulla) were segmented or when the model failed to segment either the spinal canal or spinal cord region completely. In such cases, the specific axial slice where failure occurred was excluded.
Patient Data Collection
To test our model in the clinical setting, we then collected and stored MRI data and clinical metadata for 205 consecutive patients presenting with complaints related to cervical stenosis at our institution between 2019 and 2020. As part of their treatment course, all of these patients had been independently evaluated by 1 of 4 board-certified neurosurgeons during that same period. Patients were excluded if their imaging data could not be accessed or processed—an issue that might occur if their imaging had been obtained at an outside institution. Accessible MRIs were downloaded in DICOM format and preprocessed as described earlier. Further processing and measurements were performed using these preprocessed images.
Model Evaluation
We then allowed our model to perform automated segmentations on these MRIs using previously described trained deep learning ensemble segmentation algorithms. The corresponding biomarker generated by these automated segmentations was tested against the outcome of a subsequent surgical decompression for cervical stenosis.
Results
Baseline characteristics for the test cohort are summarized in Table 1. Of the 205 patients initially identified, 58 were excluded for having inaccessible MRI data (obtained at an outside institution), leaving 147 patient images against which to test our model. Of these patients, 67 ultimately underwent surgical decompression for cervical stenosis, while 80 were managed nonoperatively after clinical evaluation by a spine surgeon.
Comparison of age and gender among surgical and nonsurgical patients.
Across the cohort, the mean minimum difference in cross-sectional area between the spinal canal and spinal cord (ie, the proposed imaging biomarker) differed significantly between patients who ultimately underwent surgical decompression and those managed nonoperatively (P = 5.29 × 10⁻³). Surgical patients had smaller mean canal–cord area differences than nonsurgical patients (35.90 ± 25.00 mm² vs 48.55 ± 33.52 mm²), consistent with greater cord compression in the surgical cohort (Table 2). The distributions of biomarker values for each group are shown in Figure 2.
Boxplot showing the statistically significant difference in the distribution of proposed biomarkers in surgical and nonsurgical cohorts.
Comparison of the mean values of the proposed biomarker between nonsurgical and surgical cohorts using an independent 2-sample t test.
By solely relying on the radiographic biomarker generated for each automatically segmented image, our model was able to predict (based on retrospectively collected data) whether a patient would ultimately undergo surgery for decompression of their cervical stenosis with relatively high accuracy (area under the curve [AUC] 0.79). In other words, a higher degree of radiographic stenosis, as quantified by our biomarker, was associated with a greater chance of subsequent surgery.
The receiver operating characteristic analysis and the corresponding confusion matrix for our model, when tested on this cohort, are presented in Figure 3 and Table 3. The confusion matrix summarizes model performance across true positive, true negative, false positive, and false negative classifications. Of the 67 patients who ultimately underwent surgical decompression, the model correctly classified 61 as surgical (true positives) and misclassified 6 as nonsurgical (false negatives). Among the 80 patients managed nonoperatively, 61 were correctly classified as nonsurgical (true negatives), while 19 were misclassified as surgical (false positives). Misclassification occurred in both directions, reflecting overlap between surgical and nonsurgical cohorts based on imaging-derived stenosis severity alone.
Receiver operating characteristic (ROC) curve corresponding to the prediction of surgical status using the proposed biomarker.
Confusion matrix corresponding to the prediction of surgical status using the proposed biomarker.
Discussion
In this study, we describe a proof-of-concept for an automated image analysis tool capable of identifying operative cervical stenosis using only data from MRIs images. Our model distinguished between surgical and nonsurgical cervical stenosis with high accuracy (AUC 0.79) by relying on an automated biomarker representing the degree of space between the spinal cord and the surrounding spinal canal. This biomarker was constructed because it is both physiologic and intuitive; it follows that as the area of the spinal canal approaches that of the spinal cord, spinal cord compression becomes more severe. Moreover, these results support the intuitive notion that increases in compression severity correlate with a higher likelihood of surgical intervention.
The success of our model in stratifying patient surgical outcomes is worth discussing both in terms of the components it does and does not include. No model for surgical decision-making should rely solely on imaging data; however, the relatively high accuracy of our model using MRI data in isolation underscores the relative importance of radiographic features in operative triage.
Previous work by our group utilized deep learning to generate a parallel biomarker for lumbar stenosis to predict whether a patient’s lumbar pathology warranted surgical decompression.10 In comparison to our cervical model, the lumbar biomarker demonstrated a higher degree of predictive accuracy (AUC 0.88). This is partially a consequence of the inherent anatomical differences between these 2 regions, the most obvious of which is the presence of the spinal cord coursing within the cervical spinal canal. As such, in cases of cervical stenosis, variables such as cord deformity and changes in T2 signal within the cord itself can often play a role in surgical decision-making, and for the moment, these variables remain absent from our model. Ongoing work by our group aims to incorporate these components into a more comprehensive algorithm.
As a proof of concept, our model is currently not intended to assist in physicians’ clinical decision-making. Rather, the real value of the present algorithm lies in optimizing the referral process to surgery. As any spine surgeon can attest, a substantial number of patients, more than half in many cases, referred to see a spine surgeon never undergo surgery.11–13 While some of these nonoperative patients still benefit from time spent in consultation with a surgeon, the degree of inefficiency with the referral process can lead to longer wait times for patients with operative pathology and delays in care for those with nonoperative pathology (ie, delaying referral to a physiatrist or physical therapy).14,15 Spine patients are particularly vulnerable to this inefficiency due in part to the relative nuance of radiographic interpretation of spinal pathology. Given both the incidence of neck and back-related complaints during visits to primary care practitioners and the prevalence of nonoperative degenerative changes on routine MRIs of the spine, it is not surprising that so many patients arrive at the office of a spine surgeon only to be reassured that their findings do not require surgery.16–19 This problem is not novel, and various efforts have been made to improve the efficiency of referrals to spine surgeons with some success. Current efforts include the development of multidisciplinary pathways and clinical practice guidelines.20,21 However, no studies have focused on the problem of variability among referring providers with respect to discerning operative from nonoperative spinal pathology.
The goal of the present model is to provide referring physicians tasked with reviewing a patient’s cervical spine MRI with a greater degree of confidence when making the decision to refer to a spine surgeon. In its current form, the model is not intended to recommend treatment but rather to provide additional information that might meaningfully inform a referring physician’s decision to send a patient to see a spine surgeon. Future iterations of our stenosis biomarker could be used to generate a surgical likelihood score, informing referring providers of how likely a patient’s stenosis is to require surgery. To be clear, our model would not serve as a replacement for the clinical judgment of a physician but would instead help provide additional actionable information, alongside a patient’s clinical history, signs, symptoms, and physical examination findings, with which to make the decision to refer to a specialist. This additional information could help increase the number of patients with surgical pathology who are referred to a surgeon, as well as decrease the number referred without surgical pathology who could instead be referred for nonoperative treatment modalities.
Limitations
There are several limitations to our study. First, our model is susceptible to the domain effect since the decision regarding surgical candidacy was limited to a group of physicians at a single institution, limiting its generalizability.22 Second, as a proof of concept, our model focused on cases of cervical stenosis based on net canal-cord compromise to the exclusion of other surgical pathologies of the cervical spine; it does not address other forms of cervical pathology warranting referral to a subspecialist, such as foraminal stenosis, malignancy, trauma, or other disease processes.23,24 Furthermore, our model does not capture those patients without surgical pathology for whom surgical referral is warranted for reasons beyond clinical history or physical examination, such as for reassurance or to gather more information about their condition, or who did not undergo surgery for reasons other than the severity of their radiographic findings. Such cases go unnoticed by our model but are important cases for referral. Nevertheless, the findings demonstrate that imaging-derived features alone can meaningfully stratify patients by likelihood of undergoing surgical intervention, underscoring the contribution of radiographic data in referral and triage contexts rather than serving as a substitute for comprehensive clinical evaluation. Last, the data analyzed within the present study were derived from 2019 to 2020; although utilization of more recent data would allow for increased generalizability across evolving scanner hardware and imaging workflows, protocols for cervical spine imaging and radiographic features of cervical stenosis have remained stable over this period.25 These limitations notwithstanding, the results of the present study demonstrate that imaging-derived data from cervical spine MRI enable differentiation between surgical and nonsurgical patient cohorts based on the degree of cervical stenosis.
Conclusion
The present study describes a novel, fully autonomous deep learning algorithm that uses MRI-derived measures of cervical stenosis to distinguish between surgical and nonsurgical patient cohorts retrospectively. We propose that this model, and its future iterations, could serve as a valuable adjunctive tool for referring providers to consult when evaluating the need for surgical consultation in patients with cervical stenosis. This marks an initial step toward harnessing artificial intelligence to support clinical decision-making and streamlining the referral process to spine surgery.
Acknowledgments
The authors dedicate this work in honor of Dr Bilwaj Gaonkar, PhD, who passed away during the course of this study. Dr Gaonkar, a co-author of this manuscript, made a substantial contribution to the research described. The present study would not have been possible without him.
Footnotes
Funding This research was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under Award Number R21EB026665.
Disclosures Dr. Macyszyn discloses a financial interest in Theseus Artificial Intelligence, which has signed a license agreement with the University of California, Los Angeles for the technology utilized in this research. The remaining authors have no personal, financial, or institutional interest in any part of the research presented in this work.
- This manuscript is generously published free of charge by ISASS, the International Society for the Advancement of Spine Surgery. Copyright © 2026 ISASS. To see more or order reprints or permissions, see http://ijssurgery.com.
References
- 1.↵Bot SDM , van der Waal JM , Terwee CB , et al . Incidence and prevalence of complaints of the neck and upper extremity in general practice. Ann Rheum Dis. 2005;64(1):118–123. 10.1136/ard.2003.019349
- 2.↵Picavet HSJ , Schouten JSAG . Musculoskeletal pain in the Netherlands: prevalences, consequences and risk groups, the DMC3-study. Pain. 2003;102(1–2):167–178. 10.1016/s0304-3959(02)00372-x
- 3.↵Izzo R , Popolizio T , Balzano RF , Pennelli AM , Simeone A , Muto M . Imaging of cervical spine traumas. Eur J Radiol. 2019;117:75–88. 10.1016/j.ejrad.2019.05.007
- 4.↵Malhotra A , Wu X , Kalra VB , et al . Utility of MRI for cervical spine clearance after blunt traumatic injury: a meta-analysis. Eur Radiol. 2017;27(3):1148–1160. 10.1007/s00330-016-4426-z
- 5.↵Laiwalla AN , Ratnaparkhi A , Zarrin D , et al . Lumbar spinal canal segmentation in cases with lumbar stenosis using deep-U-net ensembles. World Neurosurg. 2023;178:e135–e140. 10.1016/j.wneu.2023.07.009
- 6.↵Lessmann N , van Ginneken B , de Jong PA , Išgum I . Iterative fully convolutional neural networks for automatic vertebra segmentation and identification. Med Image Anal. 2019;53:142–155. 10.1016/j.media.2019.02.005
- 7.↵Cheng P , Yang Y , Yu H , He Y . Automatic vertebrae localization and segmentation in CT with a two-stage Dense-U-Net. Sci Rep. 2021;11(1):22156. 10.1038/s41598-021-01296-1
- 8.↵Adida S , Legarreta A , Hudson JS , et al . Application of machine learning for automatic segmentation of paraspinal musculature. World Neurosurg. 2023;180:228–230. 10.1016/j.wneu.2023.10.056
- 9.↵World Medical Association Declaration of Helsinki. JAMA. 2013;310(20):2191. 10.1001/jama.2013.281053
- 10.↵Wilson B , Gaonkar B , Yoo B , et al . Predicting spinal surgery candidacy from imaging data using machine learning. Neurosurgery. 2021;89(1):116–121. 10.1093/neuros/nyab085
- 11.↵Debono B , Sabatier P , Koudsie A , Buffenoir K , Hamel O . Managing spine surgery referrals: the consultation of neurosurgery and its nuances. Neurochirurgie. 2017;63(4):267–272. 10.1016/j.neuchi.2017.05.003
- 12.↵Kidane B , Gandhi R , Sarro A , Valiante TA , Harvey BJ , Rampersaud YR . Is referral to a spine surgeon a double-edged sword? Patient concerns before consultation. Can Fam Physician. 2011;57(7):803–810.
- 13.↵Mayman D , Yen D . Maximizing use of a surgical clinic for referrals of patients having back problems. Can J Surg. 1999;42(2):117–119.
- 14.↵Findlay JM , Deis N . Appropriateness of lumbar spine referrals to a neurosurgical service. Can J Neurol Sci. 2010;37(6):843–848. 10.1017/s0317167100051544
- 15.↵Braybrooke J , Ahn H , Gallant A , et al . The impact of surgical wait time on patient-based outcomes in posterior lumbar spinal surgery. Eur Spine J. 2007;16(11):1832–1839. 10.1007/s00586-007-0452-5
- 16.↵Childress MA , Stuek SJ . Neck pain: initial evaluation and management. Am Fam Physician. 2020;102(3):150–156.
- 17.↵Bartholomeeusen S , Zundert J , Truyers C , Buntinx F , Paulus D . Higher incidence of common diagnoses in patients with low back pain in primary care. Pain Pract. 2012;12(1):1–6. 10.1111/j.1533-2500.2011.00466.x
- 18.↵Jensen MC , Brant-Zawadzki MN , Obuchowski N , Modic MT , Malkasian D , Ross JS . Magnetic resonance imaging of the lumbar spine in people without back pain. N Engl J Med. 1994;331(2):69–73. 10.1056/NEJM199407143310201
- 19.↵Kasch R , Truthmann J , Hancock MJ , et al . Association of lumbar MRI findings with current and future back pain in a population-based cohort study. Spine (Phila Pa 1986). 1976;47(3):201–211. 10.1097/brs.0000000000004198
- 20.↵Wilgenbusch CS , Wu AS , Fourney DR . Triage of spine surgery referrals through a multidisciplinary care pathway: a value-based comparison with conventional referral processes. Spine (Phila Pa 1986). 2014;39(22 suppl 1):S129–S135. 10.1097/BRS.0000000000000574
- 21.↵Layne EI , Roffey DM , Coyle MJ , Phan P , Kingwell SP , Wai EK . Activities performed and treatments conducted before consultation with a spine surgeon: are patients and clinicians following evidence-based clinical practice guidelines? Spine J. 2018;18(4):614–619. 10.1016/j.spinee.2017.08.259
- 22.↵Ratnaparkhi A , Gaonkar B , Zarrin D , et al . Ensembling mitigates scanner effects in deep learning medical image segmentation with deep-u-nets. Iftekharuddin KM , Drukker K , Mazurowski MA , Lu H , Muramatsu C , Samala RK , eds. Medical Imaging 2022: Computer-Aided Diagnosis. San Diego, California: SPIE; 2022. 10.1117/12.2611609
- 23.↵Fernández de Rota JJ , Meschian S , Fernández de Rota A , Urbano V , Baron M . Cervical spondylotic myelopathy due to chronic compression: the role of signal intensity changes in magnetic resonance images. J Neurosurg Spine. 2007;6(1):17–22. 10.3171/spi.2007.6.1.4
- 24.↵Yolas C , Ozdemir NG , Okay HO , et al . Cervical disc hernia operations through posterior laminoforaminotomy. J Craniovertebr Junction Spine. 2016;7(2):91–95. 10.4103/0974-8237.181854
- 25.↵Expert Panel on Neurological Imaging, Eldaya RW , Parsons MS , et al . ACR Appropriateness Criteria® cervical pain or cervical radiculopathy: 2024 update. J Am Coll Radiol. 2025;22(5S):S136–S162. 10.1016/j.jacr.2025.02.035
