Artificialintelligence (AI) has the potential to significantly impact the field of medical imaging by providing more efficient and accurate diagnostic tools. This potential is supported by recent publications by Anand et al. and Celtikci,1,2highlighting the potential of deep learning to automate image analysis and improve diagnostic efficacy and accuracy.
One of the conditions that may particularly benefit from the application of AI is cervical ossification of the posterior longitudinal ligament (C-OPLL), a multifactorial condition caused by abnormal lamellar bone deposition within the PLL that may lead to neurological deficits and increase perioperative complications in anterior cervical spine approaches.3–8Currently, CT is considered the diagnostic gold standard for OPLL. Despite well-validated CT-based radiographic criteria proposed by the Investigation Committee on OPLL of the Japanese Ministry of Health and Welfare,9–11diagnostic criteria for OPLL using MRI have not yet been established.12The objective of this study was to develop AI software and a validated model for the characterization and diagnosis of C-OPLL on MRI, obviating the need for spine CT.
开云体育世界杯赔率
We performed a retrospective evaluation of imaging studies in all adult patients who underwent both cervical CT and MRI for all clinical indications within a span of 36 months (between January 2017 and July 2020) in Sheba Medical Center, a single tertiary referral hospital located in central Israel. MRI was obtained by 3-T machines (Signa HD 3.0-T, GE Healthcare, and Ingenia 3.0-T, Philips). CT was acquired using a Revolution CT scanner (GE Healthcare). Patient demographics are presented inTable 1. As CT images are currently considered the gold standard for OPLL diagnosis, C-OPLL was identified using sagittally reconstructed cervical CT demonstrating the presence of diffuse, calcified, thickened PLL. C-OPLL was identified by a neurosurgery senior resident (G.K.) and reviewed by a senior neuroradiologist (G.Y.) and a senior spine neurosurgeon (R.H.;Fig. 1). C-OPLL location and maximal thickness were recorded. Basic patient demographics and the clinical indication for imaging were obtained from the MRI referral form. MATLAB software (The MathWorks, Inc.) was then used to create an AI tool for the diagnosis of C-OPLL by using a convolutional neural network model to identify features on MR images (Fig. 2). All C-OPLL carriers identified by the AI model were reviewed by two senior neurosurgeons (R.H. and a retired neurosurgeon). A reader study was performed to compare the performance of the AI model to that of the diagnostic panel using standard test performance metrics such as sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV). Overall accuracy was defined as: (true positive + true negative)/(overall number of patients). Interobserver variability and reliability were assessed using Cohen’s kappa score.
Patient demographics
| Characteristic | Negative OPLL | Positive OPLL |
|---|---|---|
| No. of patients | 735 | 65 |
| Mean age ± SD, yrs | 63.38 ± 9.2 | 62.76 ± 9.49 |
| Females, n (%) | 582 (79.2) | 45 (69.2) |
| Mean max C-OPLL thickness ± SD, mm | 2.4 ± 2.75 | 5.1 ± 1.24 |
| Clinical indications, n (%)* | ||
| Cervicalgia | 545 (74.1) | 44 (67.7) |
| Myelopathy | 301 (41) | 16 (24.6) |
| Radiculopathy | 124 (16.9) | 13 (20) |
Obtained from MRI referral form.
Study flowchart.
Artist’s illustration depicting the VGG16 convolutional neural network workflow.
Image Processing
The MATLAB medical image processing toolbox (MATLAB version 2022b, The MathWorks, Inc.) was used for image processing. The MRI DICOM files were imported and analyzed in both sagittal and axial planes using T2-weighted sequences. In addition to the use of a randomized sample of patients, data augmentation techniques were employed to increase the variability in the training data set, thereby improving the generalizability of the AI model.
The OPLL and additional anatomical structures were segmented using a semiautomatic threshold-based method, which involves setting a threshold value to separate the ossified tissue from the surrounding environment.13The segmented images were then reviewed by a senior neurosurgeon (R.H.) to ensure accuracy and to make any necessary adjustments. A set of features was extracted from the segmented images including the area, perimeter, voxel, and thickness of the OPLL. These features were used as input for the AI algorithm. Additionally, other features such as the location, shape, and intensity of the ossified tissue were extracted.
AI Algorithm
The VGG16 learning method, a pretrained convolutional neural network model that has been used extensively and effectively in the classification of images for computer vision,14was used to develop the AI algorithm. The model was fine-tuned and trained using the extracted features and the corresponding labels of the patients to improve OPLL characterization on MRI. The algorithm was implemented in MATLAB (version 2020b) with the aid of deep learning, image processing, statistics, medical image, and machine learning toolboxes. The data set was split into two sets, training (75%) and validation (25%). The data used in this study were randomized to ensure that the model was trained and validated on a representative sample of the patient population. This method of randomization minimizes any potential bias in the sample selection and increases the generalizability of the results.
Reader Study
Following the automated training and validation of the AI model, the diagnostic panel (R.H. and G.K., i.e., the "readers") independently evaluated 100 new CT and MRI scans, constituting a diagnostic benchmark for evaluating the diagnostic performance of the AI algorithm. Interobserver variability was assessed using Cohen’s kappa score to compare the accuracy of the AI algorithm to that of the reader.
Results
Nine hundred consecutive patients were found to be eligible for radiological evaluation, yielding 65 identified C-OPLL carriers. Patient demographics are displayed inTable 1. The MRI-based AI model identified C-OPLL carriers and was able to successfully segment the OPLL, vertebral bodies, and discoligamentous complex in all patients. The model identified 5 additional patients harboring C-OPLLs who were unrecognized on initial radiological screening by the diagnostic panel. A VGG16 internal architecture analysis revealed that the automated layers with the greatest impact on the model’s performance were the PLL maximal thickness, variations in signal intensity within the PLL, and the presence of kyphotic alignment. The reader study comprised 100 patients (Fig. 3). The performance of the MRI-based AI model resulted in a sensitivity of 85%, specificity of 98%, NPV of 98%, and PPV of 85%. The overall accuracy of the model was 98%, with a kappa score of 0.917.
AI algorithm confusion matrix. The matrix displays the number of true-positive, true-negative, false-positive, and false-negative results produced by the AI algorithm compared with the reader study.
Discussion
准确的识别C-OPLL高度导入ant when treating patients with neurological deficits attributed to compressive myelopathy. Anterior cervical approaches in the face of PLL ossification causing assimilation with the anterior cervical dura carry a high rate of inadvertent durotomy, associated with poor postoperative outcomes compared with posterior surgical approaches. Moreover, in severe cases, the OPLL mass may cause clinically symptomatic canal stenosis in areas beyond the reach of conventional discectomy, rendering this surgical approach inappropriate for certain patients.15,16Currently, the diagnostic gold standard for OPLL is spine CT, due to its ability to reliably demonstrate calcifications at the PLL. As the number of patients treated surgically for cervical myelopathy based on MRI alone continues to rise, the need for MRI-based identification of OPLL is becoming imperative. There are currently no diagnostic criteria for OPLL on MRI. The present study presents an MRI-based AI algorithm and a validated model for the diagnosis and anatomical representation of C-OPLL without complementary CT. The output provided by this model (Fig. 4) may readily affect the choice between surgical approaches, assist in surgical preplanning by differentiating regions harboring OPLL from unaffected areas, and increase the overall surgical safety by identifying OPLL on routine T2-weighted sequences performed prior to the vast majority of operations for cervical myelopathy. In addition, the layers underlying the self-training capacity of this model may be used to reveal diagnostic criteria for C-OPLL on MRI in the future.
C-OPLL carrier identified by the AI model.A:Bone-window cervical spine CT scan demonstrating OPLL at C6, C7, and T1. Of note, calcification of the posterior aspect of the C4–5 disc space was noted.B:矢t2加权MRI相同的病人demonstrating a thickened hypointense mass at the PLL at these levels.C:Automated C-OPLL segmentation performed by the AI algorithm. Notably, the system was trained to avoid misinterpretation of calcified discs as OPLL. The calcified C4–5 posterior osteophyte was not identified as C-OPLL by the novel AI model.
Model Performance
人工智能模型的评估检测的能力OPLL was performed through a reader study (n = 100), in which the model’s results were compared with the diagnostic agreement reached by the panel of two neurosurgeons and a neuroradiologist, who reviewed both CT and MRI scans for each patient. In the reader study, our model demonstrated excellent specificity of 98%, albeit with an intermediate sensitivity of 85%. This variation in performance is partly due to the limited number of identified C-OPLL carriers in the naïve population. Further training on larger data sets is expected to significantly enhance sensitivity. Additionally, the evaluation of these test performance metrics should be considered in light of the current absence of MRI-adjusted diagnostic criteria for C-OPLL, as there is no universally accepted definition for an MRI-based diagnostic benchmark. From a clinical perspective, the model’s exceptionally high NPV can be utilized to resolve ambiguous situations in which the surgical team confronts a dilemma in weighing the outcomes of anterior surgery in the presence of mild C-OPLL against the drawbacks of posterior approaches, such as the limited ability to correct kyphotic alignment.
Study Limitations
The technical limitations of the VGG16 convolutional neural network lie in its high computational requirements, which are not typically met by the hardware infrastructure present in healthcare facilities. This requirement could potentially hinder its widespread implementation in resource-limited healthcare environments. Furthermore, the validation set and reader study were based on CT serving as the diagnostic gold standard. This approach may result in the underestimation of the incidence of occult C-OPLL detected only by MRI. This limitation could potentially lead to an increased rate of false-negative results. Finally, the limited sample size of C-OPLL carriers decreased the sensitivity of the AI model. Upon reevaluation of discordant readings by the review panel, it was found that the majority of false-positive readings were due to the presence of ossified bulging discs. This finding highlights the need for larger sample sizes in future studies to enhance the model’s sensitivity in this common radiological finding.
Future Directions in AI-Based Diagnostics of Spinal Pathologies
The present study leverages the use of VGG16, a pretrained convolutional neural network model, as the basis for the AI algorithm. This approach allows the transfer of knowledge acquired from other data sets to enhance the performance of the algorithm. VGG16 has gained widespread use in medical imaging and has been demonstrated to be effective in various applications such as lesion detection, image segmentation, and classification.17,18人工智能模型的自动分割输出先生的images provides a foundation for the detection and representation of additional spinal pathologies, offering the benefits of CT scans without the associated risks of radiation exposure.
Conclusions
The novel AI software developed in this study showed promising results in its ability to identify C-OPLL on MRI without the use of CT scans. This promise has the potential to reduce radiation exposure for patients and support surgical preplanning and decision-making regarding the surgical approach. With further development, this MRI-based AI model has the potential to aid in the diagnosis of various spinal disorders, and its automated layers may lay the foundation for MRI-specific diagnostic criteria for C-OPLL.
Disclosures
The authors report no conflict of interest concerning the materials or methods used in this study or the findings specified in this paper.
Author Contributions
Conception and design: Harel, Shemesh, Kimchi. Acquisition of data: Shemesh, Kimchi, Yaniv. Analysis and interpretation of data: Harel, Shemesh, Kimchi. Drafting the article: Harel, Shemesh, Kimchi. Critically revising the article: Harel, Shemesh, Yaniv. Reviewed submitted version of manuscript: all authors. Approved the final version of the manuscript on behalf of all authors: Harel. Statistical analysis: Shemesh. Administrative/technical/material support: Shemesh, Yaniv. Study supervision: Harel, Shemesh.
References
-
1 ↑
AnandA,FloresAR,McDonaldMF,GadotR,XuDS,RopperAE.A computer vision approach to identifying the manufacturer of posterior thoracolumbar instrumentation systems.J Neurosurg Spine.2022;38(4):417–424.
-
2 ↑
CeltikciE.A systematic review on machine learning in neurosurgery: the future of decision-making in patient care.Turk Neurosurg.2018;28(2):167–173.
-
3 ↑
TanakaM,KanazawaA,YonenobuK.Diagnosis of OPLL and OYL: Overview. In:YonenobuK,NakamuraK,ToyamaY, eds.OPLL.Springer;2006:111–113.
-
4
BoodyBS,LendnerM,VaccaroAR.Ossification of the posterior longitudinal ligament in the cervical spine: a review.Int Orthop.2019;43(4):797–805.
-
5
KwokSSS,CheungJPY.Surgical decision-making for ossification of the posterior longitudinal ligament versus other types of degenerative cervical myelopathy: anterior versus posterior approaches.BMC Musculoskelet Disord.2020;21(1):823.
-
6
NamDC,LeeHJ,LeeCJ,HwangSC.Molecular pathophysiology of ossification of the posterior longitudinal ligament (OPLL).Biomol Ther (Seoul).2019;27(4):342–348.
-
7
FujiyoshiT,YamazakiM,KawabeJ,et al.A new concept for making decisions regarding the surgical approach for cervical ossification of the posterior longitudinal ligament: the K-line.Spine (Phila Pa 1976).2008;33(26):E990–E993.
-
9 ↑
SaetiaK,ChoD,LeeS,KimDH,KimSD.Ossification of the posterior longitudinal ligament: a review.Neurosurg Focus.2011;30(3):E1.
-
10
FiroozniaH,RafiiM,GolimbuC,TylerI,BenjaminVM,PintoRS.Computed tomography of calcification and ossification of posterior longitudinal ligament of the spine.J Comput Tomogr.1984;8(4):317–324.
-
11 ↑
NagataK,SatoK.Diagnostic imaging of cervical ossification of the posterior longitudinal ligament. In:YonenobuK,NakamuraK,ToyamaY, eds.OPLL.Springer;2006:127–143.
-
12 ↑
KangMS,LeeJW,ZhangHY,ChoYE,ParkYM.Diagnosis of cervical OPLL in lateral radiograph and MRI: is it reliable?Korean J Spine.2012;9(3):205–208.
-
13 ↑
FriedliL,KloukosD,KanavakisG,HalazonetisD,GkantidisN.The effect of threshold level on bone segmentation of cranial base structures from CT and CBCT images.Sci Rep.2020;10(1):7361.
-
14 ↑
ZhaoD,ZhuD,LuJ,LuoY,ZhangG.Synthetic medical images using F&BGAN for improved lung nodules classification by multi-scale VGG16.Symmetry.2018;10(10):519.
-
15 ↑
WuJC,ChenYC,HuangWC.Ossification of the posterior longitudinal ligament in cervical spine: prevalence, management, and prognosis.Neurospine.2018;15(1):33–41.
-
16 ↑
WuJC,LiuL,ChenYC,HuangWC,ChenTJ,ChengH.Ossification of the posterior longitudinal ligament in the cervical spine: an 11-year comprehensive national epidemiology study.Neurosurg Focus.2011;30(3):E5.
-
17 ↑
JiangZP,LiuYY,ShaoZE,HuangKW.An improved VGG16 model for pneumonia image classification.Appl Sci.2021;11(23):11185.
-
18 ↑
AlshammariA.Construction of VGG16 Convolution Neural Network (VGG16_CNN) classifier with NestNet-based segmentation paradigm for brain metastasis classification.Sensors (Basel).2022;22(20):8076.
