External Validation of SpineNetv2 Deep Learning System for Automated Lumbar Spine MRI Analysis: A Multi-pathology Diagnostic Accuracy Study

External Validation of SpineNetv2 Deep Learning System for Automated Lumbar Spine MRI Analysis: A Multi-pathology Diagnostic Accuracy Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article External Validation of SpineNetv2 Deep Learning System for Automated Lumbar Spine MRI Analysis: A Multi-pathology Diagnostic Accuracy Study Xinkai Wu, Qianbo Song, Jiaxiang Zhou, Zhiyu Zhou, Guangru Cao, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7559680/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 08 Nov, 2025 Read the published version in European Spine Journal → Version 1 posted 10 You are reading this latest preprint version Abstract Background Magnetic resonance imaging (MRI) is the reference standard for evaluating degenerative lumbar spine disorders, but interpretation is time-consuming and subject to inter-observer variability. SpineNetv2, a publicly available deep learning system, enables automated analysis of multiple spinal pathologies. This study conducted an independent external validation of SpineNetv2 against expert reference standards. Methods A total of 491 patients (2,455 lumbar discs, L1/2–L5/S1) were retrospectively included. Disc-level reference labels were established by an expert orthopedic surgeon, with a junior orthopedic surgeon serving as comparator. Six pathologies were assessed: disc degeneration (Pfirrmann grading), central canal stenosis (CCS), spondylolisthesis, herniation, and bilateral foraminal stenosis (FS). Performance metrics included sensitivity, specificity, positive predictive value, negative predictive value, F1-score, Matthews correlation coefficient, exact agreement, weighted kappa, and mean absolute error. McNemar’s test and bootstrap resampling (1,000 iterations) were used for statistical analysis. Results Overall accuracy ranged from 83.5–97.5% (mean 92.8%). SpineNetv2 significantly outperformed the junior orthopedic surgeon in CCS, spondylolisthesis, and bilateral FS (all p ≤ 0.001), with comparable performance in herniation (p = 0.293). Pfirrmann grading showed lower MAE for SpineNetv2 compared with the junior surgeon (0.213 vs. 0.254, p = 0.001), though accuracy declined in older patients and upper lumbar discs. Error analysis revealed a specificity-oriented profile, with false negatives exceeding false positives. Conclusions SpineNetv2 demonstrated high accuracy across five binary lumbar pathologies, while Pfirrmann grading remained the main limitation, particularly in elderly upper lumbar discs. Its specificity-oriented profile supports use as a confirmatory second reader, but reliance on negative findings is not recommended. Broader reliability will require multicenter, multi-reader validation and sensitivity-oriented calibration. Lumbar spine Magnetic resonance imaging Degenerative pathology SpineNetv2 Artificial intelligence Diagnostic performance External validation Pfirrmann grading Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 INTRODUCTION Low back pain is the leading cause of disability worldwide and has consistently ranked as the primary contributor to years lived with disability over the past three decades[ 1 ]. Epidemiological evidence indicates that 70–85% of adults experience at least one episode of low back pain during their lifetime[ 2 ]. Among its various causes, degenerative changes in the lumbar spine, including intervertebral disc degeneration, disc herniation, spinal stenosis, and spondylolisthesis, are the most common structural contributors[ 3 ]. These conditions not only reduce mobility and quality of life but also generate considerable healthcare and socioeconomic burdens. Magnetic resonance imaging (MRI) is the gold standard for evaluating lumbar spine pathology, offering high-resolution assessment of disc morphology, hydration status, and neural element compression[ 4 , 5 ]. MRI-based classification systems, such as the Pfirrmann disc degeneration grading scale, are widely applied in both clinical and research settings[ 6 ]. Nevertheless, MRI interpretation remains complex and subjective[ 7 ]. Diagnostic accuracy depends heavily on the reader’s expertise, with studies reporting only moderate inter-observer agreement even among experienced clinicians[ 8 ]. In addition, the growing demand for spinal MRI examinations places increasing pressure on radiology services, making timely and consistent interpretation more difficult[ 9 ]. With the rapid development of artificial intelligence (AI) and deep learning, automated image analysis has become a promising tool in musculoskeletal imaging[ 10 ]. Convolutional neural networks and related architectures have shown strong performance in various spinal imaging tasks, including grading intervertebral disc degeneration, detecting disc herniation, and identifying spinal stenosis[ 11 , 12 ]. Several studies have reported diagnostic accuracies comparable to those of radiologists, suggesting that AI systems may improve clinical efficiency and reduce human error in spine MRI interpretation[ 8 , 12 ]. Despite these advances, important limitations continue to restrict the clinical utility of existing approaches. Many prior studies have focused on a single disease or diagnostic task, limiting their relevance to the complex and multifaceted presentations of lumbar spine pathology[ 13 ]. Furthermore, most algorithms have been developed and validated on relatively small, single-center datasets, raising concerns about robustness and generalizability to independent cohorts[ 14 ]. In addition, the reliability of reference standards remains problematic, as many earlier evaluations relied on a single reader, introducing subjectivity and potential bias[ 14 , 15 ]. To address these limitations, we performed an external validation of SpineNetv2, a publicly available system for automated detection and grading of lumbar spine MRI, using a dataset that was independent of model development[ 16 ]. Based on expert-established reference standards, we systematically compared the diagnostic performance of SpineNetv2 with that of both a junior and an expert orthopedic surgeon across multiple common lumbar pathologies. We further analyzed diagnostic concordance, error patterns, and performance stability across patient subgroups. This independent validation framework was designed to quantify the diagnostic reliability of SpineNetv2, identify considerations for clinical deployment, and provide evidence-based recommendations for AI-assisted lumbar spine diagnosis. MATERIALS and METHODS Study design and population This retrospective diagnostic accuracy study was conducted in accordance with the STARD guidelines. Consecutive lumbar spine MRI examinations performed at our institution between January 2023 and December 2024 were analyzed. The study protocol was approved by the Institutional Review Board (IRB No. KYLL-2025-0012). Inclusion criteria were adult patients (≥ 18 years) who underwent lumbar spine MRI with complete imaging from L1–L2 through L5–S1 disc levels. Exclusion criteria included prior spinal surgery, incomplete imaging sequences, or poor image quality that precluded diagnostic interpretation. Reference standards and evaluation protocol Disc-level reference labels were established by an expert orthopedic surgeon with more than 20 years of experience in lumbar spine disorders and MRI interpretation. All gradings were performed using standardized protocols and were blinded to both SpineNetv2 outputs and junior surgeon assessments. To ensure the reliability of the reference standard, the expert reviewer re-evaluated a random subset of 200 cases after a 4-week washout period, achieving high intra-rater agreement (κ = 0.89–0.94 across diseases). A junior orthopedic surgeon independently assessed all cases using the same protocols, serving as a performance comparator but not contributing to the establishment of reference standards. SpineNetv2 system and disease evaluation framework SpineNetv2 is a publicly available deep learning framework developed by the Visual Geometry Group at the University of Oxford for automated analysis of lumbar spine MRI[ 16 , 17 ]. The system performs automated vertebral detection, segmental landmark identification, and pathological grading using a convolutional neural network architecture optimized for multi-pathology spinal assessment. Six spinal pathologies were evaluated using established clinical classification frameworks, with the original grading criteria summarized in Table 1 . Table 1 Comprehensive disease classification criteria Grading System Description Pfirrmann 1 Homogeneous disc, hyperintense, normal height 2 Inhomogeneous disc, hyperintense, normal height 3 Inhomogeneous disc, isointense, normal/decreased height 4 Inhomogeneous disc, hypointense, normal/decreased height 5 Inhomogeneous disc, hypointense, collapsed disc Central Canal Stenosis 1 Normal 2 Mild, compromise ≤ 1⁄3 of normal size 3 Moderate, compromise 1⁄3–2⁄3 of normal size 4 Severe, compromise of > 2⁄3 of normal size Spondylolisthesis 0 Normal, ≤ 2 mm slippage, or < 25% slippage (Meyerding I) 1 Mild, 25–50% slippage (Meyerding II) 2 Moderate, 51–75% slippage (Meyerding III) 3 Severe (Meyerding IV) Disc Herniation 0 Normal disc morphology, no focal protrusion or extrusion 1 Disc herniation present Bilateral Foraminal Stenosis 0 Normal neural foramen, uncompromised nerve root passage 1 Foraminal stenosis present, neural foraminal narrowing with potential nerve root compromise During preliminary analysis, categories with a prevalence of < 5% or fewer than 100 cases were identified for potential consolidation[ 18 ]. Consolidation was guided by evidence-based clinical principles while preserving the distinction between normal and pathological states. After consolidation, binary classification thresholds were applied: Pfirrmann disc degeneration (≥ Grade 4), central canal stenosis (CCS, ≥Grade 2 after merging severe grades), and spondylolisthesis (≥ Grade 1 after merging higher grades). Disc herniation and bilateral foraminal stenosis (FS) were inherently binary outcomes (present vs. absent)[ 19 – 21 ]. Outcome measures The primary outcome was diagnostic concordance between SpineNetv2 and expert reference standards across six spinal pathologies. For binary classifications, the primary metrics were sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Secondary metrics included the F1-score and Matthews correlation coefficient (MCC). For the multi-level Pfirrmann grading, the primary outcomes were exact agreement rates and weighted kappa coefficients, with mean absolute error (MAE) as a secondary measure. Error characterization was performed using confusion matrix analysis and severity quantification, distinguishing minor (± 1 grade) from major ( ≥ ± 2 grades) discrepancies. Comparative performance against the junior orthopedic surgeon served as an additional benchmark across all pathologies. Statistical analysis Patient-level bootstrap resampling (1,000 iterations) was applied to derive 95% confidence intervals for all diagnostic performance metrics. Entire patients were resampled, and all disc levels were carried forward to preserve within-patient correlation. Paired comparisons between SpineNetv2 and the junior surgeon were performed using McNemar’s test at the patient level, with disease presence defined as a positive prediction at any of the five lumbar levels. Prevalence gradients across spinal levels were evaluated using the Cochran–Armitage trend test. Paired ordinal outcomes (Pfirrmann grading) were compared using the Wilcoxon signed-rank test. Determinants of diagnostic accuracy were assessed using disc-level regression models with patient-clustered robust standard errors to account for within-patient correlation. Logistic regression was applied to binary disease outcomes, and ordinal logistic regression to multi-level disease grading. Fixed effects included age group ( 60 years), sex, spinal segment (upper L1–L3 vs. lower L4–S1), and disease category, where applicable. Multiple comparisons were controlled using the Benjamini–Hochberg false discovery rate (q = 0.05) within prespecified families: primary diagnostic performance comparisons and subgroup analyses within diseases. All statistical tests were two-sided (α = 0.05). Analyses were performed in Python 3.11.7 with NumPy (1.26.4), SciPy (1.13.1), statsmodels (0.14.2), and scikit-learn (1.5.1). RESULTS Dataset characteristics and baseline demographics The final cohort included 491 patients, providing 2,455 intervertebral disc levels (five per patient: L1–L2, L2–L3, L3–L4, L4–L5, and L5–S1), all of which were evaluated by the expert orthopedic surgeon, the junior orthopedic surgeon, and SpineNetv2. Age distribution deviated from normality (Shapiro–Wilk test, p < 0.05) ( Supplementary Fig. 1 ). Six spinal pathologies were assessed: three with ordinal grading systems (disc degeneration, CCS, and spondylolisthesis) and three with binary classification (disc herniation and bilateral FS). All MRI examinations were performed according to standard clinical protocols for routine lumbar spine evaluation. Detailed baseline characteristics are summarized in Table 2 . Table 2 Patient baseline characteristics and dataset structure (n = 491) Characteristics Value Total Patients 491 Age, years Median (IQR) 40.0 (31.0, 50.0) Range 17–89 Age groups, n (%) 60 years 53 (10.8%) Sex, n (%) M 301 (61.3%) F 190 (38.7%) Dataset Structure Total disc levels analyzed 2,455 Spinal segments per patient 5 Data completeness 100% Disease classification processing and prevalence analysis Systematic evaluation of disease grade distributions indicated that CCS and spondylolisthesis required grade consolidation due to rare categories (< 5% prevalence or < 100 cases). For CCS, Grades 3 and 4 were merged into Grade 2, yielding a binary classification of Grade 1 versus Grades 2–4. For spondylolisthesis, Grades 2 and 3 were merged into Grade 1, resulting in a binary classification of Grade 0 versus Grades 1–3 ( Supplementary Fig. 2 ). The final analytical framework therefore consisted of one multilevel disease (Pfirrmann grading) and five binary diseases. Disease prevalence analysis showed patient-level prevalence ranging from 25.7% for spondylolisthesis to 55.4% for herniation. The Cochran–Armitage trend test demonstrated significant cranial-to-caudal increasing gradients for Pfirrmann grading, herniation, and both FS conditions (all p < 0.001). Complete prevalence patterns and segmental distributions are illustrated in Fig. 1 . Diagnostic performance evaluation Diagnostic performance was assessed across five diseases analyzed as binary classifications and Pfirrmann grading evaluated as a multi-level ordinal classification. SpineNetv2 demonstrated superior overall performance compared with the junior orthopedic surgeon in most pathologies. SpineNetv2 achieved significantly higher diagnostic accuracy than the junior orthopedic surgeon for CCS (p = 0.001), spondylolisthesis (p < 0.001), and bilateral FS (p < 0.001). Performance for herniation was comparable between the two approaches (p = 0.293). SpineNetv2 consistently achieved higher specificity and positive predictive values, whereas the junior orthopedic surgeon demonstrated greater sensitivity across most conditions. For Pfirrmann grading, SpineNetv2 showed superior diagnostic concordance, with a significantly lower mean absolute error (0.213 vs. 0.254, p = 0.001) compared with the junior orthopedic surgeon (Fig. 2 , Fig. 3 ). Error pattern characterization Systematic evaluation of SpineNetv2 error patterns across six spinal pathologies showed overall diagnostic accuracy ranging from 83.5–97.5%, with a mean accuracy of 92.8%. For binary classifications, error directionality was consistent, with false negatives substantially exceeding false positives: CCS (5.4% vs. 0.4%), herniation (5.0% vs. 1.5%), and bilateral FS (4.7–5.9% vs. 0.2–0.5%), reflecting the model’s conservative diagnostic profile (Fig. 4 ). For Pfirrmann grading, the most common misclassification involved Grade 2 cases predicted as Grade 1 (106 cases, representing 19.3% of all Grade 2 discs). Mild errors (± 1 grade) accounted for 11.6% of cases, whereas severe errors ( ≥ ± 2 grades) occurred in only 4.9% (Fig. 5 ). Risk factors for diagnostic discordance Comprehensive risk stratification across 36 demographic–spinal combinations (six diseases × three age groups × two spinal levels) showed predominantly reliable performance of SpineNetv2. Overall diagnostic consistency averaged 92.7% across all subgroups, with risk distribution as follows: 30 low-risk (83.3%), five medium-risk (13.9%), and one high-risk (2.8%) combinations. Risk concentration was disease-specific, with disc degeneration accounting for all medium- and high-risk scenarios. The only high-risk combination was observed in older patients (> 60 years) with upper lumbar disc degeneration (77.7% consistency, 94 cases). The five medium-risk combinations corresponded to the remaining Pfirrmann age–segment strata (81.6–86.9% consistency). All other diseases demonstrated consistently high reliability across demographic and segmental subgroups, ranging from 90.8% for CCS to 97.9% for spondylolisthesis (Fig. 6 ). Multivariable performance determinants analysis Following disease consolidation, multivariable analysis was conducted using disc-level logistic regression for binary diseases (12,275 observations) and ordinal logistic regression for Pfirrmann grading (2,455 observations), both with patient-clustered robust standard errors. Significant associations were observed in 5 of 24 factor–disease combinations (20.8%), with overall accuracy averaging 94.9% across subgroups. Age effects were limited to Pfirrmann grading, with reduced accuracy in middle-aged (OR = 0.54, 95% CI: 0.35–0.84, p = 0.007) and older patients (OR = 0.27, 95% CI: 0.15–0.46, p < 0.001). Male sex was associated with improved accuracy for CCS (OR = 1.61, p = 0.006) and herniation (OR = 1.45, p = 0.024). Lower lumbar segments outperformed upper segments only in CCS (OR = 1.50, p = 0.020). No other factor–disease combinations demonstrated significant associations (Fig. 7 ). DISCUSSION In this independent external validation study, we evaluated the performance of SpineNetv2 in 491 patients, encompassing 2,455 lumbar discs across six common degenerative pathologies. The model achieved high diagnostic accuracy, ranging from 83.5–97.5% (mean 92.8%), and consistently outperformed the junior orthopedic surgeon in most binary tasks. Accuracy was significantly higher for CCS, spondylolisthesis, and bilateral FS, whereas performance for disc herniation was comparable. For Pfirrmann grading, SpineNetv2 showed superior concordance with a significantly lower MAE (0.213 vs. 0.254), although accuracy declined in older patients and at upper lumbar segments. Taken together, these findings demonstrate a diagnostic profile characterized by high specificity and positive predictive values but relatively lower sensitivity, reflecting a conservative algorithmic tendency that prioritizes minimizing false positives. Early deep learning models primarily focused on single diagnostic tasks, with weighted kappa values for Pfirrmann grading ranging from 0.59 to 0.87 [ 22 ]. Recent external validations of SpineNet reported class-balanced accuracies of 74–79% and kappa values of 0.63–0.77[ 23 – 25 ]. The higher accuracy observed in our study (average 92.8%) may be explained by several factors. First, we consolidated rare disease grades, thereby simplifying complex multi-class problems into binary tasks. While this likely improved apparent performance, it also reduced granularity in severity assessment. Second, we used overall accuracy as the primary metric rather than class-balanced accuracy, which may overestimate performance under imbalanced disease prevalence. Third, SpineNetv2 may incorporate algorithmic optimizations beyond the original SpineNet. Most importantly, its conservative diagnostic profile, characterized by high specificity and low sensitivity, reduces false positives but risks missing clinically relevant cases. Thus, despite favorable statistical metrics, it is essential to consider how these methodological factors influence the real-world diagnostic utility of the system. Detailed error analysis revealed a consistent conservative diagnostic pattern across all binary classifications, with false negatives substantially outnumbering false positives. This pattern reflects deliberate specificity prioritization rather than a limitation in capability and is likely driven by several mechanisms. First, class imbalance in the training data, combined with standard decision thresholds, naturally suppresses false positives. Second, the inherent ambiguity of borderline classifications, particularly at transition zones such as Pfirrmann Grades 1–2 or CCS Grades 1–2, predisposes the algorithm to underdiagnose uncertain cases. This explains why the most frequent misclassification involved Pfirrmann Grade 2 cases predicted as Grade 1, with most errors limited to ± 1 grade. Third, common imaging confounders, including motion artifacts, endplate sclerosis, Modic changes, and anatomical variations such as short pedicles in the upper lumbar spine, can obscure subtle pathological features, further contributing to conservative predictions[ 26 ]. This diagnostic profile contrasts with the tendency of junior clinicians to over-diagnose, accounting for SpineNetv2’s superior positive predictive values despite lower sensitivity. From a clinical perspective, these findings suggest that deployment strategies should prioritize threshold recalibration to improve sensitivity without materially reducing specificity. In practice, the system appears best suited as a confirmatory second reader, particularly valuable for ruling in pathology, whereas negative predictions should not be used as the sole basis for reassurance in high-risk patients or complex cases[ 23 , 27 ]. Risk stratification indicated that nearly all medium- to high-risk scenarios were confined to the Pfirrmann task, while other endpoints remained stable across demographic and segmental strata. The only high-risk subgroup, elderly patients with upper lumbar discs, showed 77.7% accuracy, highlighting the intrinsic limitations of automated ordinal degeneration assessment. In advanced age, composite degenerative changes compress the T2 dynamic range and blur ordinal boundaries between Grades 3–5. At upper lumbar levels, thinner discs and anatomical crowding increase partial-volume effects and reduce conspicuity. Reader uncertainty is correspondingly higher in these near-boundary states, which is propagated to the model. Technical factors further exacerbate this vulnerability. Pfirrmann grading relies on subtle intensity differences, while disc-level context and segment-wise domain shifts can weaken positive evidence. Representation of this subgroup in our cohort was limited, which may hinder calibration and threshold optimization, regardless of the undisclosed training distribution. In contrast, binary endpoints maintained high accuracy across strata, suggesting that presence/absence decisions are less susceptible to these confounders. These findings support the need for age- and segment-conditioned calibration, ordinal-aware training objectives with enriched sampling of elderly upper-lumbar cases, and mandatory expert review for this predefined high-risk group. This study has several limitations that should be considered in interpretation. The single-center, single-protocol cohort and reliance on a single-expert reference standard, despite high intra-rater agreement (κ = 0.89–0.94), limit generalizability compared with multi-reader, multi-site designs. Consolidation of rare severe grades stabilized estimation but reduced ordinal granularity and likely inflated apparent accuracy, a notable constraint for pre-surgical stratification where differentiating severe from very severe disease guides treatment decisions. Class imbalance and cohort-specific prevalence further complicate cross-study comparisons, particularly against balanced metrics or alternative endpoints. The absence of clinical correlation restricts conclusions to imaging-based accuracy rather than downstream clinical impact. Sparse data in key strata, most notably elderly upper-lumbar cases, produced imprecise estimates and warrant cautious interpretation, underscoring the need for prospective multicenter studies with multi-reader standards, ordinal endpoints, and sensitivity-oriented calibration. Future validation should prioritize multicenter studies with consensus reference standards to confirm SpineNetv2’s specificity-oriented diagnostic profile across diverse populations and imaging protocols[ 27 ]. In the near term, technical refinements should focus on age- and segment-conditioned calibration for Pfirrmann grading and on assessing whether modest threshold relaxation can improve sensitivity without materially reducing specificity, particularly in the high-risk elderly upper-lumbar subgroup[ 26 ]. Beyond algorithmic adjustments, prospective evaluations are needed to determine whether operating at a high-specificity threshold provides tangible clinical benefits, including fewer unnecessary interventions, greater workflow efficiency, and improved diagnostic confidence, when implemented as a second reader. Pending such evidence, our findings support a limited but important role for SpineNetv2 as a confirmatory tool for positive findings rather than a stand-alone diagnostic system[ 7 ]. CONCLUSION SpineNetv2 demonstrated high accuracy across five binary lumbar pathologies, while Pfirrmann grading remained the main limitation, particularly in elderly upper lumbar discs. Its specificity-oriented profile supports use as a confirmatory second reader, but reliance on negative findings is not recommended. Broader reliability will require multicenter, multi-reader validation and sensitivity-oriented calibration. Declarations Ethics approval This study has been approved by the Ethics Committee of our hospital (IRB number KYLL-2025-0012). All procedures involving human participants are conducted in accordance with institutional guidelines, the Helsinki Declaration, and relevant national regulations. Consent to participate All patients provided written informed consent for the use of their clinical and imaging data for research purposes. Consent was obtained at the time of hospital admission or outpatient consultation using standard institutional consent forms approved by the Ethics Committee. Patients who declined to provide consent were excluded from the study. Conflicts of interest All authors declare that they have no conflict of interest. Availability of data and material The MRI image data used in this study were obtained from patients at the Second Affiliated Hospital of Zunyi Medical University. Due to patient privacy and institutional regulations, the raw data cannot be made publicly available. Data may be accessed upon reasonable request and with approval from the Institutional Review Board of the Second Affiliated Hospital of Zunyi Medical University. Acknowledgements I express my sincere gratitude for the unwavering support from my family members. Author contribution Conceptualization: Kebing Jin, Qian Du. Data curation: Xingkai Wu, Qianbo Song. Formal Analysis: Jiaxiang Zhou, Zhiyu Zhou, Guangru Cao. Funding acquisition: Zhiyu Zhou, Kebing Jin, Qian Du. Investigation/ Supervision: Zhiyu Zhou, Guangru Cao, Xingkai Wu, Qianbo Song, Jiaxiang Zhou, Kebing Jin, Qian Du. Methodology: Xingkai Wu, Jiaxiang Zhou, Kebing Jin. Project administration: Guangru Cao, Ke bing Jin, Qian Du. Resources: Guangru Cao, Qian Du. Writing - original draft: Xingkai Wu. Writing - review & editing: Qianbo Song, Jiaxaing Zhou, Zhiyu Zhou, Guangru Cao, Kebing Jin, Qian Du. Xingkai Wu and Qianbo Song contributed equally to this work. Funding This research was funded by the Basic Research Program of Guizhou Provincial Department of Science and Technology (Qiankehe Foundation-ZK [2024] General-347), the Science and Technology Plan Projects of Zunyi City (Zunshi Kehe HZ (2024) No. 432 and 442), and Guizhou Science and Technology Plan Project (No. QKHRC-KJZY[2025]036). References Hoy D, March L, Brooks P, Blyth F, et al. (2014) The global burden of low back pain: estimates from the Global Burden of Disease 2010 study.Ann Rheum Dis 73(6):968-974. https://doi.org/10.1136/annrheumdis-2013-204428 de Souza I, Sakaguchi TF, Yuan S, Matsutani LA, et al. (2019) Prevalence of low back pain in the elderly population: a systematic review.Clinics (Sao Paulo) 74(e789. https://doi.org/10.6061/clinics/2019/e789 Knezevic NN, Candido KD, Vlaeyen J, Van Zundert J, Cohen SP (2021) Low back pain.Lancet 398(10294):78-92. https://doi.org/10.1016/S0140-6736(21)00733-9 Beaudart C, McCloskey E, Bruyère O, Cesari M, et al. (2016) Sarcopenia in daily practice: assessment and management.BMC Geriatr 16(1):170. https://doi.org/10.1186/s12877-016-0349-4 Liang YW, Fang YT, Lin TC, Yang CR, et al. (2024) The Quantitative Evaluation of Automatic Segmentation in Lumbar Magnetic Resonance Images.Neurospine 21(2):665-675. https://doi.org/10.14245/ns.2448060.030 Xie J, Yang Y, Jiang Z, Zhang K, et al. (2023) MRI radiomics-based decision support tool for a personalized classification of cervical disc degeneration: a two-center study.Front Physiol 14(1281506. https://doi.org/10.3389/fphys.2023.1281506 Liawrungrueang W, Park JB, Cholamjiak W, Sarasombath P, Riew KD (2025) Artificial Intelligence-Assisted MRI Diagnosis in Lumbar Degenerative Disc Disease: A Systematic Review.Global Spine J 15(2):1405-1418. https://doi.org/10.1177/ 21925682241274372 Zheng HD, Sun YL, Kong DW, Yin MC, et al. (2022) Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI.Nat Commun 13(1):841. https://doi.org/10.1038/s41467-022-28387-5 Chen X, Wang X, Zhang K, Fung KM, et al. (2022) Recent advances and clinical applications of deep learning in medical image analysis.Med Image Anal 79(102444. https://doi.org/10.1016/j.media.2022.102444 Tsai JY, Hung IY, Guo YL, Jan YK, et al. (2021) Lumbar Disc Herniation Automatic Detection in Magnetic Resonance Imaging Based on Deep Learning.Front Bioeng Biotechnol 9(708137. https://doi.org/10.3389/fbioe.2021.708137 Beulah A, Sharmila TS, Pramod VK (2021) Degenerative disc disease diagnosis from lumbar MR images using hybrid features.The Visual Computer 38(8): https://doi.org/10. 1007/ s00371-021-02154-x Lehnen NC, Haase R, Faber J, Rüber T, et al. (2021) Detection of Degenerative Changes on MR Images of the Lumbar Spine with a Convolutional Neural Network: A Feasibility Study.Diagnostics (Basel) 11(5):902. https://doi.org/10.3390/diagnostics11050902 Wang ZX, Hu YG (2012) High-intensity zone (HIZ) of lumbar intervertebral disc on T2-weighted magnetic resonance images: spatial distribution, and correlation of distribution with low back pain (LBP).Eur Spine J 21(7):1311-1315. https://doi.org/10.1007/ s00586-012-2240-0 Yi W, Zhao J, Tang W, Yin H, et al. (2023) Deep learning-based high-accuracy detection for lumbar and cervical degenerative disease on T2-weighted MR images.Eur Spine J 32(11):3807-3814. https://doi.org/10.1007/s00586-023-07641-4 Abdollah V, Parent EC, Battié MC (2019) Reliability and validity of lumbar disc height quantification methods using magnetic resonance images.Biomed Tech (Berl) 64(1):111-117. https://doi.org/10.1515/bmt-2017-0086 Windsor, R., Jamaludin, A., Kadir, T, et al. (2022) SpineNetV2: Automated Detection, Labelling and Radiological Grading Of Clinical MR Scans. ArXiv, abs/2205.01683. https://doi.org/10.48550/arXiv.2205.01683 Windsor R, Jamaludin A, Kadir T, Zisserman A (2024) Automated detection, labelling and radiological grading of clinical spinal MRIs.Sci Rep 14(1):14993. https://doi.org/10.1038/s41598-024-64580-w Burke JF, Sussman JB, Kent DM, Hayward RA (2015) Three simple rules to ensure reasonably credible subgroup analyses.BMJ 351(h5651. https://doi.org/10.1136/bmj.h5651 Chen XL, Li XY, Wang Y, Lu SB (2023) Relation of lumbar intervertebral disc height and severity of disc degeneration based on Pfirrmann scores.Heliyon 9(10):e20764. https://doi.org/10.1016/j.heliyon.2023.e20764 Andreisek G, Imhof M, Wertli M, Winklhofer S, et al. (2013) A systematic review of semiquantitative and qualitative radiologic criteria for the diagnosis of lumbar spinal stenosis.AJR Am J Roentgenol 201(5):W735-746. https://doi.org/10.2214/AJR.12.10163 Koslosky E, Gendelberg D (2020) Classification in Brief: The Meyerding Classification System of Spondylolisthesis.Clin Orthop Relat Res 478(5):1125-1130. https://doi.org/10.1097/CORR.0000000000001153 Baur D, Bieck R, Berger J, Schöfer P, et al. (2025) Automated Three-Dimensional Imaging and Pfirrmann Classification of Intervertebral Disc Using a Graphical Neural Network in Sagittal Magnetic Resonance Imaging of the Lumbar Spine.J Imaging Inform Med 38(2):979-987. https://doi.org/10.1007/s10278-024-01251-2 Grob A, Loibl M, Jamaludin A, Winklhofer S, et al. (2022) External validation of the deep learning system "SpineNet" for grading radiological features of degeneration on MRIs of the lumbar spine.Eur Spine J 31(8):2137-2148. https://doi.org/10.1007/s00586-022-07311-x McSweeney TP, Tiulpin A, Saarakkala S, Niinimäki J, et al. (2023) External Validation of SpineNet, an Open-Source Deep Learning Model for Grading Lumbar Disk Degeneration MRI Features, Using the Northern Finland Birth Cohort 1966.Spine (Phila Pa 1976) 48(7):484-491. https://doi.org/10.1097/BRS.0000000000004572 Nigru AS, Benini S, Bonetti M, Bragaglio G, et al. (2024) External validation of SpineNetV2 on a comprehensive set of radiological features for grading lumbosacral disc pathologies.N Am Spine Soc J 20(100564. https://doi.org/10.1016/j.xnsj.2024.100564 Nikpasand M, Middendorf JM, Ella VA, Jones KE, et al. (2024) Automated magnetic resonance imaging-based grading of the lumbar intervertebral disc and facet joints.JOR Spine 7(3):e1353. https://doi.org/10.1002/jsp2.1353 Zhao P, Zhu S (2025) Advances and challenges in AI-assisted MRI for lumbar disc degeneration detection and classification.Eur Spine J. https://doi.org/10.1007/s00586-025-09179-z Additional Declarations No competing interests reported. Supplementary Files SupplementaryFigure1.png Fig. S1 Patient demographics. Left: histogram of patient age for the study cohort (n=491); dashed lines mark the mean (41.5 years, red) and median (40.0 years, orange). Right: sex distribution pie chart (male 61.3%, female 38.7%). Values shown in the panels are patient-level SupplementaryFigure2.png Fig. S2 Disc-level grade distributions and sample adequacy. (A) Pfirrmann disc degeneration (grades 1–5); (B) central canal stenosis (CCS; grades 1–4); (C) spondylolisthesis (grades 0–2). Bars display the proportion of discs per grade with counts annotated. Percentages are calculated over all evaluable discs (L1/2–L5/S1). Strata flagged as under-represented did not meet the a priori sample-adequacy criterion (Methods) and were not analyzed at native granularity; per protocol, CCS grades 3–4 were pooled within grade ≥2 and spondylolisthesis grade 2 within grade ≥1. Pfirrmann was retained as an ordinal endpoint. Color coding denotes sample adequacy (adequate vs under-represented). Abbreviation: CCS, central canal stenosis SupplementaryFigure1.png Fig. S1 Patient demographics. Left: histogram of patient age for the study cohort (n=491); dashed lines mark the mean (41.5 years, red) and median (40.0 years, orange). Right: sex distribution pie chart (male 61.3%, female 38.7%). Values shown in the panels are patient-level Cite Share Download PDF Status: Published Journal Publication published 08 Nov, 2025 Read the published version in European Spine Journal → Version 1 posted Editorial decision: Revision requested 08 Oct, 2025 Reviews received at journal 06 Oct, 2025 Reviews received at journal 25 Sep, 2025 Reviewers agreed at journal 25 Sep, 2025 Reviewers agreed at journal 23 Sep, 2025 Reviewers agreed at journal 10 Sep, 2025 Reviewers invited by journal 10 Sep, 2025 Editor assigned by journal 09 Sep, 2025 Submission checks completed at journal 09 Sep, 2025 First submitted to journal 07 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7559680","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":514212788,"identity":"4e0f225d-7994-4830-a5af-d5449e0addda","order_by":0,"name":"Xinkai Wu","email":"","orcid":"","institution":"The Second Affiliated Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xinkai","middleName":"","lastName":"Wu","suffix":""},{"id":514212789,"identity":"c36840eb-ee4d-4bd0-929b-ae7a740cfb16","order_by":1,"name":"Qianbo Song","email":"","orcid":"","institution":"The Second Affiliated Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Qianbo","middleName":"","lastName":"Song","suffix":""},{"id":514212790,"identity":"a04901d8-a886-4e28-9377-9ebec720d602","order_by":2,"name":"Jiaxiang Zhou","email":"","orcid":"","institution":"The Seventh Affiliated Hospital of Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Jiaxiang","middleName":"","lastName":"Zhou","suffix":""},{"id":514212791,"identity":"0eadb1f7-c5e7-421b-909b-93a912f45df4","order_by":3,"name":"Zhiyu Zhou","email":"","orcid":"","institution":"The Seventh Affiliated Hospital of Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Zhiyu","middleName":"","lastName":"Zhou","suffix":""},{"id":514212795,"identity":"44bde0f1-2332-4e03-88a6-626a0043ef38","order_by":4,"name":"Guangru Cao","email":"","orcid":"","institution":"The Second Affiliated Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Guangru","middleName":"","lastName":"Cao","suffix":""},{"id":514212798,"identity":"d9eaf5a8-a20b-401b-abb3-76af035c1ee4","order_by":5,"name":"Kebing Jin","email":"","orcid":"","institution":"Guizhou University","correspondingAuthor":false,"prefix":"","firstName":"Kebing","middleName":"","lastName":"Jin","suffix":""},{"id":514212800,"identity":"9a2b92d1-73eb-433f-8003-96dfbdd77780","order_by":6,"name":"Qian Du","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvUlEQVRIiWNgGAWjYDCCA2BkA+HwkKAljUQtQHCYBC18N5I3Hvi547y97owExgdv2xjkzQlpkbyRVnCw98xtZrMbCcyGc9sYDHc2ENBicCPH4ABv2202oBY2ad42hgSDA0RoOfi37RwPUAv7b6K1HOZtOyABsoWZKC2SZ54VHJZtSzYwO/OwWXLOOQnDDYS08B1P3vzxbZudvdnx5IMf3pTZyBO0BeQ2KM3YACQkCKtH0jIKRsEoGAWjAAcAABiURGQMux/cAAAAAElFTkSuQmCC","orcid":"","institution":"The Second Affiliated Hospital of Zunyi Medical University","correspondingAuthor":true,"prefix":"","firstName":"Qian","middleName":"","lastName":"Du","suffix":""}],"badges":[],"createdAt":"2025-09-08 03:53:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7559680/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7559680/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00586-025-09543-z","type":"published","date":"2025-11-08T15:57:42+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":91828464,"identity":"379bc5d5-9d5d-44f5-ae89-15d9a813110e","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"tif","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1129848,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1DiseasePrevalenceOverviewFixed.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/94b307f82351d446ce299b3e.tif"},{"id":91828493,"identity":"0fb6f57e-172f-4444-8a99-19d467111c47","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6374510,"visible":true,"origin":"","legend":"","description":"","filename":"AnonymisedManuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/434405f4a282f44e6e5af550.docx"},{"id":91828458,"identity":"04c10dbe-293c-4927-ad11-6a0b26b86545","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"tif","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2434600,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2EnhancedBinaryPerformanceCI.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/930011ee437a896364f0d46a.tif"},{"id":91828489,"identity":"a6741491-2607-4f30-9fe6-6d07f564471f","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":165413,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3EnhancedMultilevelPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/f7d994e22f54bef6e4f30e7e.png"},{"id":91828430,"identity":"739ce9a9-3a58-418f-952d-1ff4e440d46b","added_by":"auto","created_at":"2025-09-22 08:53:28","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":19333,"visible":true,"origin":"","legend":"","description":"","filename":"Table1Comprehensivediseaseclassificationcriteria.docx","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/fea19835fb940b251958bbae.docx"},{"id":91828474,"identity":"685f4693-06c6-46e5-8af0-4cb759883798","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":435648,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4CompleteAIConfusionMatrices.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/8baf20843da448744d74c248.png"},{"id":91828497,"identity":"80be1b90-d6d9-4037-ad6e-db6b911931e7","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18383,"visible":true,"origin":"","legend":"","description":"","filename":"Table2Detailedbaselinecharacteristics.docx","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/3d66039d9b15e9b89b79fe80.docx"},{"id":91828470,"identity":"814e8fcb-72d7-4849-abdd-f2f4f686403c","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"tif","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":853896,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5AIErrorAnalysis.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/98b3b5612b8da4d1b4266296.tif"},{"id":91830305,"identity":"2c77eed3-30d9-4df6-afd9-0014635087a0","added_by":"auto","created_at":"2025-09-22 09:01:29","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":156878,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/8d21c05acb6194863c141b47.png"},{"id":91828503,"identity":"e1ae8335-e36c-4c4f-891f-5a18e9b71c97","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":235282,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/5d903751991fd9d1b99f7a0c.png"},{"id":91828479,"identity":"c18add33-5322-44d4-9a8c-e328882fffa0","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"json","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8998,"visible":true,"origin":"","legend":"","description":"","filename":"ee7752b7e61340638b987576de069931.json","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/4fb0bedb03d530cea00c4994.json"},{"id":91828496,"identity":"25348401-04cf-481a-8718-0b76697d6b0c","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"docx","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16905,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/c8e67b5bc55e0d6da66db96c.docx"},{"id":91828498,"identity":"b8227bfd-d7db-4a25-a5c6-698ddb1062d0","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"xml","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104281,"visible":true,"origin":"","legend":"","description":"","filename":"ee7752b7e61340638b987576de0699311enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/fa19f653fc7abdadd421fb3f.xml"},{"id":91828476,"identity":"892580e6-ec43-4fa2-9ee9-71e8d8e59548","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"tif","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1129848,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1DiseasePrevalenceOverviewFixed.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/562908a6e6a9c266a0c9ef56.tif"},{"id":91830310,"identity":"ab458466-17e0-4d69-bb90-9a8a4190027d","added_by":"auto","created_at":"2025-09-22 09:01:32","extension":"tif","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2434600,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2EnhancedBinaryPerformanceCI.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/537917ae6722d970445e22d4.tif"},{"id":91828520,"identity":"d00b9dcd-c199-48a0-93ee-31e2c407b309","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":165413,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3EnhancedMultilevelPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/7f0aff01026641172cf4391a.png"},{"id":91830308,"identity":"680071bc-3e7f-4bcf-acca-1ecc96d9597e","added_by":"auto","created_at":"2025-09-22 09:01:31","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":435648,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4CompleteAIConfusionMatrices.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/3a6b0f576a5a7dd21ea5acdf.png"},{"id":91830311,"identity":"dd2c36dc-e8aa-4c6d-a55c-fc2428415d8e","added_by":"auto","created_at":"2025-09-22 09:01:32","extension":"tif","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":853896,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5AIErrorAnalysis.tif","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/8f1ae16c65aef5076a31b737.tif"},{"id":91828494,"identity":"4bf6ad6b-75e8-430c-87fa-f89238f84607","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":262144,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6AIConsistencyRiskMatrix.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/8f1f04a69172a75fd6877e4e.png"},{"id":91828508,"identity":"299caeda-b49f-4f92-8d53-4701cc357a3e","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":262144,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6AIConsistencyRiskMatrix.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/543064258762808cff95e29b.png"},{"id":91828490,"identity":"d3c2a0c7-e861-44b6-8af1-d2b76cb47e93","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"jpeg","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1129848,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1DiseasePrevalenceOverviewFixed.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/2e2f4aa344d47296a162cd28.jpeg"},{"id":91828486,"identity":"b0abd756-542a-48c2-9a88-f270155992eb","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"jpeg","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2434600,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2EnhancedBinaryPerformanceCI.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/ec70a8998bc24f4713be7bd9.jpeg"},{"id":91830313,"identity":"ce9e11ea-49c7-4a21-bc6f-0c955cb98e2d","added_by":"auto","created_at":"2025-09-22 09:01:32","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":165413,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3EnhancedMultilevelPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/69700ceb94fdfdb5721dbe92.png"},{"id":91828511,"identity":"d5e0925e-a3ee-4274-a8c4-52bc2b86fe41","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":435648,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4CompleteAIConfusionMatrices.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/c47d5be450c5b8065d9c4421.png"},{"id":91828488,"identity":"eb2ba792-b9b6-4dd1-8b84-86f2206959e4","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"jpeg","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":853896,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5AIErrorAnalysis.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/534e1a373ac7a065a52f8a2b.jpeg"},{"id":91828466,"identity":"7348844e-fd2f-40ef-8c25-894da83aba24","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":431540,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/6fd218b8c716f70ca8b9292c.png"},{"id":91828492,"identity":"18867f65-7137-4a05-ae67-1d3a82cc45dd","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":466724,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/2d7e6801e5efe38d306bb1d1.png"},{"id":91830314,"identity":"54fb1edf-3b5e-447f-b030-56da55b8a444","added_by":"auto","created_at":"2025-09-22 09:01:32","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":705918,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1DiseasePrevalenceOverviewFixed.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/ecb430895967c6da7bc1bbcd.png"},{"id":91828530,"identity":"f1ff1571-97ad-48b7-8fff-96517be6aab8","added_by":"auto","created_at":"2025-09-22 08:53:33","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1407886,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2EnhancedBinaryPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/358b72e859fc21b54675055c.png"},{"id":91828483,"identity":"62ff8ca0-cdbc-45b6-9bbb-ec9091b1a1c6","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":109793,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3EnhancedMultilevelPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/56a4f6ac8d9975a332b46b3e.png"},{"id":91828536,"identity":"7bc4be38-11e7-47b0-825f-17dd90ef2c1a","added_by":"auto","created_at":"2025-09-22 08:53:34","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":204572,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4CompleteAIConfusionMatrices.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/321a04b85ecb447f0bc3a80f.png"},{"id":91828533,"identity":"093288c1-edf1-4ba6-a972-2fb2e5a4d63a","added_by":"auto","created_at":"2025-09-22 08:53:33","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":528815,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5AIErrorAnalysis.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/05648252f867c0f14a0b934c.png"},{"id":91828505,"identity":"e9e50658-7386-4f1b-bc8d-18df261ebf7c","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":63997,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/9f42fb26b4771d93a1a395e6.png"},{"id":91828525,"identity":"f7c6e425-3591-48d8-b5b6-adb7f4192a4a","added_by":"auto","created_at":"2025-09-22 08:53:33","extension":"png","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":145125,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineSupplementaryFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/be5d8cf2060ceeb1ba0522f3.png"},{"id":91828499,"identity":"4be37521-e6ef-4b1c-83c4-71c1b2a13435","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"png","order_by":40,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":63997,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/78f01c2f9f213140f11dd778.png"},{"id":91828529,"identity":"016c55be-1a2e-41cd-9738-b16b86b40a6f","added_by":"auto","created_at":"2025-09-22 08:53:33","extension":"png","order_by":41,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":145125,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/572f38c1977401f3cb9efd11.png"},{"id":91828514,"identity":"4abc2fa4-4b56-4765-b6f9-2b7e250fb68b","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":42,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":705918,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/830665e84ee4719383219512.png"},{"id":91828480,"identity":"e687259c-76f6-49aa-9b9b-7a057319f789","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"png","order_by":43,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1407886,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/a64d256804e9a343fcb4526d.png"},{"id":91830309,"identity":"b9e14dab-0215-40b3-afd4-e0ea68b32f69","added_by":"auto","created_at":"2025-09-22 09:01:31","extension":"png","order_by":44,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":109793,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/6fabde76f4a4f065250f27df.png"},{"id":91828523,"identity":"5e213d9b-afbd-4b99-ac6d-ed2bb0e07a46","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":45,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":204572,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/476f8452b03c223c373feeaf.png"},{"id":91830312,"identity":"24ba04eb-6e27-4d0e-be6d-564569becf25","added_by":"auto","created_at":"2025-09-22 09:01:32","extension":"png","order_by":46,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":528815,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/af10bef74c898ae18faee75a.png"},{"id":91828504,"identity":"6598827c-e356-4f0a-8266-7444c10474d6","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":47,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":216413,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/1681c493b474c4118ebf9125.png"},{"id":91828501,"identity":"61100c4f-e22f-44d2-9127-5df8cc6d5205","added_by":"auto","created_at":"2025-09-22 08:53:31","extension":"png","order_by":48,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":212361,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/75ab17de2dc162373a38b600.png"},{"id":91828521,"identity":"f395bbdf-b7e9-4e3d-951e-098ec9688f59","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"xml","order_by":49,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98932,"visible":true,"origin":"","legend":"","description":"","filename":"ee7752b7e61340638b987576de0699311structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/871fb9f1b49ec1ce97728f85.xml"},{"id":91828506,"identity":"2e2f7404-706b-4318-ab26-bb7d5bfb715f","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"html","order_by":50,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111646,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/024050a27019d27d8142cf57.html"},{"id":91828484,"identity":"21a916a2-10f0-4ed2-8cea-7556cb4b8415","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1075631,"visible":true,"origin":"","legend":"\u003cp\u003ePrevalence and segment-wise trends across six lumbar pathologies. (A) Disc-level versus patient-level prevalence. All diseases are displayed as binary endpoints following consolidation procedures: Pfirrmann grading (≥Grade 4 vs \u0026lt;Grade 4), central canal stenosis (≥Grade 2 vs Grade 1), spondylolisthesis (≥Grade 1 vs Grade 0), disc herniation and bilateral foraminal stenosis (present vs absent). (B) Segment-specific prevalence from L1-L2 to L5-S1 with linear trend testing (Cochran-Armitage). Arrows indicate gradient direction, with upward arrows representing increasing trends. ***p \u0026lt; 0.001, ns = not significant. Abbreviation: CCS, central canal stenosis; FS, foraminal stenosis\u003c/p\u003e","description":"","filename":"Figure1DiseasePrevalenceOverviewFixed.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/ad26c05ac75168667533e6a6.png"},{"id":91828467,"identity":"948fbbc1-5037-4b5d-b88f-5fc254c012f4","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2790622,"visible":true,"origin":"","legend":"\u003cp\u003eBinary classification performance comparison showing sensitivity, specificity, positive and negative predictive values, F1-scores, and Matthews correlation coefficients with 95% confidence intervals from patient-level bootstrap resampling. * indicates significant difference (McNemar test, p \u0026lt; 0.05). Abbreviation: CCS, central canal stenosis; FS, foraminal stenosis\u003c/p\u003e","description":"","filename":"Figure2EnhancedBinaryPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/aab946c8d9089913fd00846d.png"},{"id":91828471,"identity":"a43cfd44-1266-4de5-84d9-698ebad724f8","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":615113,"visible":true,"origin":"","legend":"\u003cp\u003eMulti-level classification performance analysis for disc degeneration displaying exact agreement rates, weighted kappa coefficients, and mean absolute error with 95% confidence intervals\u003c/p\u003e","description":"","filename":"Figure3EnhancedMultilevelPerformanceCI.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/0f0ebe947c86b77a73154c12.png"},{"id":91828487,"identity":"ff96ab27-5294-4d1a-9f50-179a1828652f","added_by":"auto","created_at":"2025-09-22 08:53:30","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1223978,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices for all six spinal diseases showing SpineNetv2 predictions against ground truth, with percentages and case counts displayed for each classification cell. Ordinal→Binary: originally multi-level, analyzed as binary\u003c/p\u003e","description":"","filename":"Figure4CompleteAIConfusionMatrices.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/c0ba11158cf2b62fd6774d1a.png"},{"id":91830306,"identity":"61206fb9-0023-4e88-85be-94b4e14532e6","added_by":"auto","created_at":"2025-09-22 09:01:30","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":843080,"visible":true,"origin":"","legend":"\u003cp\u003eError severity distribution for multi-level diseases and error direction composition for binary diseases, displayed as stacked bar charts showing proportions of correct predictions, mild errors, and severe errors. Abbreviation: CCS, central canal stenosis; FS, foraminal stenosis\u003c/p\u003e","description":"","filename":"Figure5AIErrorAnalysis.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/5757703bdc67d47c411618d3.png"},{"id":91828440,"identity":"4a79158f-1ae5-4c66-a0bf-10f807f1ccf8","added_by":"auto","created_at":"2025-09-22 08:53:28","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":325043,"visible":true,"origin":"","legend":"\u003cp\u003eRisk assessment matrix showing SpineNetv2 diagnostic consistency across age groups and spinal levels for each disease, with color-coded consistency scores and case counts displayed in a 3×2 grid format for each pathology. Risk Level Interpretation: Consistency ≥ 0.9, Low Risk (Reliable); Consistency 0.8-0.9, Medium Risk (Caution Needed); Consistency \u0026lt; 0.8, High Risk (Manual Review Required)\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/598472642ba24e04ea9cedeb.png"},{"id":91831601,"identity":"cfbb1bb3-79dc-4a0d-bcc1-3040edb1e225","added_by":"auto","created_at":"2025-09-22 09:09:30","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":345127,"visible":true,"origin":"","legend":"\u003cp\u003eMultivariable logistic regression results displaying odds ratios and 95% confidence intervals derived from patient-clustered robust standard errors, with significant associations highlighted. OR \u0026gt; 1, factor increases SpineNetv2 accuracy; OR \u0026lt; 1, factor decreases SpineNetv2 accuracy; OR \u0026gt; 2, strong positive effect; OR \u0026lt; 0.5, strong negative effect. Red diamonds, statistically significant (p \u0026lt; 0.05); Blue circles, non-significant effects. Abbreviation: CCS, central canal stenosis; FS, foraminal stenosis\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/d3624982c4d911ed8dfa6fc2.png"},{"id":95564384,"identity":"76bc3417-c5d9-4a1a-8198-790b526015af","added_by":"auto","created_at":"2025-11-10 16:09:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":8052664,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/7cff405e-f7fe-4747-ac10-3ae6dbab33d1.pdf"},{"id":91828510,"identity":"46d4963f-e337-4399-973c-adb9f58ca700","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":156878,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFig. S1\u003c/strong\u003e Patient demographics. Left: histogram of patient age for the study cohort (n=491); dashed lines mark the mean (41.5 years, red) and median (40.0 years, orange). Right: sex distribution pie chart (male 61.3%, female 38.7%). Values shown in the panels are patient-level\u003c/p\u003e","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/2986db7ca37b4b13951da812.png"},{"id":91828459,"identity":"24e84ed2-781f-4024-abed-8938011e710a","added_by":"auto","created_at":"2025-09-22 08:53:29","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":235282,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFig. S2\u003c/strong\u003e Disc-level grade distributions and sample adequacy. (A) Pfirrmann disc degeneration (grades 1–5); (B) central canal stenosis (CCS; grades 1–4); (C) spondylolisthesis (grades 0–2). Bars display the proportion of discs per grade with counts annotated. Percentages are calculated over all evaluable discs (L1/2–L5/S1). Strata flagged as under-represented did not meet the a priori sample-adequacy criterion (Methods) and were not analyzed at native granularity; per protocol, CCS grades 3–4 were pooled within grade ≥2 and spondylolisthesis grade 2 within grade ≥1. Pfirrmann was retained as an ordinal endpoint. Color coding denotes sample adequacy (adequate vs under-represented). Abbreviation: CCS, central canal stenosis\u003c/p\u003e","description":"","filename":"SupplementaryFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/19f2247ff70589d6eb57d038.png"},{"id":91828516,"identity":"b9827626-d0c9-408c-a850-8b4e44044191","added_by":"auto","created_at":"2025-09-22 08:53:32","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"supplement","size":156878,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFig. S1\u003c/strong\u003e Patient demographics. Left: histogram of patient age for the study cohort (n=491); dashed lines mark the mean (41.5 years, red) and median (40.0 years, orange). Right: sex distribution pie chart (male 61.3%, female 38.7%). Values shown in the panels are patient-level\u003c/p\u003e","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7559680/v1/8f383ca84b4ef39374458ec4.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"External Validation of SpineNetv2 Deep Learning System for Automated Lumbar Spine MRI Analysis: A Multi-pathology Diagnostic Accuracy Study","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eLow back pain is the leading cause of disability worldwide and has consistently ranked as the primary contributor to years lived with disability over the past three decades[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Epidemiological evidence indicates that 70\u0026ndash;85% of adults experience at least one episode of low back pain during their lifetime[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Among its various causes, degenerative changes in the lumbar spine, including intervertebral disc degeneration, disc herniation, spinal stenosis, and spondylolisthesis, are the most common structural contributors[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These conditions not only reduce mobility and quality of life but also generate considerable healthcare and socioeconomic burdens.\u003c/p\u003e\u003cp\u003eMagnetic resonance imaging (MRI) is the gold standard for evaluating lumbar spine pathology, offering high-resolution assessment of disc morphology, hydration status, and neural element compression[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. MRI-based classification systems, such as the Pfirrmann disc degeneration grading scale, are widely applied in both clinical and research settings[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Nevertheless, MRI interpretation remains complex and subjective[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Diagnostic accuracy depends heavily on the reader\u0026rsquo;s expertise, with studies reporting only moderate inter-observer agreement even among experienced clinicians[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In addition, the growing demand for spinal MRI examinations places increasing pressure on radiology services, making timely and consistent interpretation more difficult[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eWith the rapid development of artificial intelligence (AI) and deep learning, automated image analysis has become a promising tool in musculoskeletal imaging[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Convolutional neural networks and related architectures have shown strong performance in various spinal imaging tasks, including grading intervertebral disc degeneration, detecting disc herniation, and identifying spinal stenosis[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Several studies have reported diagnostic accuracies comparable to those of radiologists, suggesting that AI systems may improve clinical efficiency and reduce human error in spine MRI interpretation[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDespite these advances, important limitations continue to restrict the clinical utility of existing approaches. Many prior studies have focused on a single disease or diagnostic task, limiting their relevance to the complex and multifaceted presentations of lumbar spine pathology[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Furthermore, most algorithms have been developed and validated on relatively small, single-center datasets, raising concerns about robustness and generalizability to independent cohorts[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. In addition, the reliability of reference standards remains problematic, as many earlier evaluations relied on a single reader, introducing subjectivity and potential bias[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo address these limitations, we performed an external validation of SpineNetv2, a publicly available system for automated detection and grading of lumbar spine MRI, using a dataset that was independent of model development[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Based on expert-established reference standards, we systematically compared the diagnostic performance of SpineNetv2 with that of both a junior and an expert orthopedic surgeon across multiple common lumbar pathologies. We further analyzed diagnostic concordance, error patterns, and performance stability across patient subgroups. This independent validation framework was designed to quantify the diagnostic reliability of SpineNetv2, identify considerations for clinical deployment, and provide evidence-based recommendations for AI-assisted lumbar spine diagnosis.\u003c/p\u003e"},{"header":"MATERIALS and METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy design and population\u003c/h2\u003e\u003cp\u003e This retrospective diagnostic accuracy study was conducted in accordance with the STARD guidelines. Consecutive lumbar spine MRI examinations performed at our institution between January 2023 and December 2024 were analyzed. The study protocol was approved by the Institutional Review Board (IRB No. KYLL-2025-0012).\u003c/p\u003e\u003cp\u003eInclusion criteria were adult patients (\u0026ge;\u0026thinsp;18 years) who underwent lumbar spine MRI with complete imaging from L1\u0026ndash;L2 through L5\u0026ndash;S1 disc levels. Exclusion criteria included prior spinal surgery, incomplete imaging sequences, or poor image quality that precluded diagnostic interpretation.\u003c/p\u003e\u003cp\u003e\u003cb\u003eReference standards and evaluation protocol\u003c/b\u003e\u003c/p\u003e\u003cp\u003eDisc-level reference labels were established by an expert orthopedic surgeon with more than 20 years of experience in lumbar spine disorders and MRI interpretation. All gradings were performed using standardized protocols and were blinded to both SpineNetv2 outputs and junior surgeon assessments. To ensure the reliability of the reference standard, the expert reviewer re-evaluated a random subset of 200 cases after a 4-week washout period, achieving high intra-rater agreement (κ\u0026thinsp;=\u0026thinsp;0.89\u0026ndash;0.94 across diseases). A junior orthopedic surgeon independently assessed all cases using the same protocols, serving as a performance comparator but not contributing to the establishment of reference standards.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eSpineNetv2 system and disease evaluation framework\u003c/h3\u003e\n\u003cp\u003eSpineNetv2 is a publicly available deep learning framework developed by the Visual Geometry Group at the University of Oxford for automated analysis of lumbar spine MRI[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The system performs automated vertebral detection, segmental landmark identification, and pathological grading using a convolutional neural network architecture optimized for multi-pathology spinal assessment. Six spinal pathologies were evaluated using established clinical classification frameworks, with the original grading criteria summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComprehensive disease classification criteria\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eGrading System\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDescription\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"4\" rowspan=\"5\"\u003e\u003cp\u003ePfirrmann\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHomogeneous disc, hyperintense, normal height\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInhomogeneous disc, hyperintense, normal height\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInhomogeneous disc, isointense, normal/decreased height\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInhomogeneous disc, hypointense, normal/decreased height\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInhomogeneous disc, hypointense, collapsed disc\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eCentral Canal Stenosis\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNormal\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMild, compromise\u0026thinsp;\u0026le;\u0026thinsp;1\u0026frasl;3 of normal size\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModerate, compromise 1\u0026frasl;3\u0026ndash;2\u0026frasl;3 of normal size\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSevere, compromise of \u0026gt;\u0026thinsp;2\u0026frasl;3 of normal size\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eSpondylolisthesis\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNormal, \u0026le; 2 mm slippage, or \u0026lt;\u0026thinsp;25% slippage (Meyerding I)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMild, 25\u0026ndash;50% slippage (Meyerding II)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModerate, 51\u0026ndash;75% slippage (Meyerding III)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSevere (Meyerding IV)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eDisc Herniation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNormal disc morphology, no focal protrusion or extrusion\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDisc herniation present\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eBilateral Foraminal Stenosis\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNormal neural foramen, uncompromised nerve root passage\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eForaminal stenosis present, neural foraminal narrowing with potential nerve root compromise\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eDuring preliminary analysis, categories with a prevalence of \u0026lt;\u0026thinsp;5% or fewer than 100 cases were identified for potential consolidation[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Consolidation was guided by evidence-based clinical principles while preserving the distinction between normal and pathological states. After consolidation, binary classification thresholds were applied: Pfirrmann disc degeneration (\u0026ge;\u0026thinsp;Grade 4), central canal stenosis (CCS, \u0026ge;Grade 2 after merging severe grades), and spondylolisthesis (\u0026ge;\u0026thinsp;Grade 1 after merging higher grades). Disc herniation and bilateral foraminal stenosis (FS) were inherently binary outcomes (present vs. absent)[\u003cspan additionalcitationids=\"CR20\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eOutcome measures\u003c/h3\u003e\n\u003cp\u003eThe primary outcome was diagnostic concordance between SpineNetv2 and expert reference standards across six spinal pathologies. For binary classifications, the primary metrics were sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Secondary metrics included the F1-score and Matthews correlation coefficient (MCC). For the multi-level Pfirrmann grading, the primary outcomes were exact agreement rates and weighted kappa coefficients, with mean absolute error (MAE) as a secondary measure. Error characterization was performed using confusion matrix analysis and severity quantification, distinguishing minor (\u0026plusmn;\u0026thinsp;1 grade) from major (\u0026thinsp;\u0026ge;\u0026thinsp;\u0026plusmn;\u0026thinsp;2 grades) discrepancies. Comparative performance against the junior orthopedic surgeon served as an additional benchmark across all pathologies.\u003c/p\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003ePatient-level bootstrap resampling (1,000 iterations) was applied to derive 95% confidence intervals for all diagnostic performance metrics. Entire patients were resampled, and all disc levels were carried forward to preserve within-patient correlation. Paired comparisons between SpineNetv2 and the junior surgeon were performed using McNemar\u0026rsquo;s test at the patient level, with disease presence defined as a positive prediction at any of the five lumbar levels.\u003c/p\u003e\u003cp\u003ePrevalence gradients across spinal levels were evaluated using the Cochran\u0026ndash;Armitage trend test. Paired ordinal outcomes (Pfirrmann grading) were compared using the Wilcoxon signed-rank test.\u003c/p\u003e\u003cp\u003eDeterminants of diagnostic accuracy were assessed using disc-level regression models with patient-clustered robust standard errors to account for within-patient correlation. Logistic regression was applied to binary disease outcomes, and ordinal logistic regression to multi-level disease grading. Fixed effects included age group (\u0026lt;\u0026thinsp;40, 40\u0026ndash;60, \u0026gt;\u0026thinsp;60 years), sex, spinal segment (upper L1\u0026ndash;L3 vs. lower L4\u0026ndash;S1), and disease category, where applicable.\u003c/p\u003e\u003cp\u003eMultiple comparisons were controlled using the Benjamini\u0026ndash;Hochberg false discovery rate (q\u0026thinsp;=\u0026thinsp;0.05) within prespecified families: primary diagnostic performance comparisons and subgroup analyses within diseases. All statistical tests were two-sided (α\u0026thinsp;=\u0026thinsp;0.05). Analyses were performed in Python 3.11.7 with NumPy (1.26.4), SciPy (1.13.1), statsmodels (0.14.2), and scikit-learn (1.5.1).\u003c/p\u003e\u003c/div\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eDataset characteristics and baseline demographics\u003c/h2\u003e\u003cp\u003eThe final cohort included 491 patients, providing 2,455 intervertebral disc levels (five per patient: L1\u0026ndash;L2, L2\u0026ndash;L3, L3\u0026ndash;L4, L4\u0026ndash;L5, and L5\u0026ndash;S1), all of which were evaluated by the expert orthopedic surgeon, the junior orthopedic surgeon, and SpineNetv2. Age distribution deviated from normality (Shapiro\u0026ndash;Wilk test, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (\u003cb\u003eSupplementary Fig.\u0026nbsp;1\u003c/b\u003e). Six spinal pathologies were assessed: three with ordinal grading systems (disc degeneration, CCS, and spondylolisthesis) and three with binary classification (disc herniation and bilateral FS). All MRI examinations were performed according to standard clinical protocols for routine lumbar spine evaluation. Detailed baseline characteristics are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePatient baseline characteristics and dataset structure (n\u0026thinsp;=\u0026thinsp;491)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCharacteristics\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal Patients\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e491\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge, years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMedian (IQR)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e40.0 (31.0, 50.0)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRange\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e17\u0026ndash;89\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge groups, n (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026lt;40 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e244 (49.7%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e40\u0026ndash;60 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e194 (39.5%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026gt;60 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e53 (10.8%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSex, n (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e301 (61.3%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e190 (38.7%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset Structure\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal disc levels analyzed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2,455\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpinal segments per patient\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eData completeness\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e100%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eDisease classification processing and prevalence analysis\u003c/h3\u003e\n\u003cp\u003eSystematic evaluation of disease grade distributions indicated that CCS and spondylolisthesis required grade consolidation due to rare categories (\u0026lt;\u0026thinsp;5% prevalence or \u0026lt;\u0026thinsp;100 cases). For CCS, Grades 3 and 4 were merged into Grade 2, yielding a binary classification of Grade 1 versus Grades 2\u0026ndash;4. For spondylolisthesis, Grades 2 and 3 were merged into Grade 1, resulting in a binary classification of Grade 0 versus Grades 1\u0026ndash;3 (\u003cb\u003eSupplementary Fig.\u0026nbsp;2\u003c/b\u003e). The final analytical framework therefore consisted of one multilevel disease (Pfirrmann grading) and five binary diseases.\u003c/p\u003e\u003cp\u003eDisease prevalence analysis showed patient-level prevalence ranging from 25.7% for spondylolisthesis to 55.4% for herniation. The Cochran\u0026ndash;Armitage trend test demonstrated significant cranial-to-caudal increasing gradients for Pfirrmann grading, herniation, and both FS conditions (all p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Complete prevalence patterns and segmental distributions are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eDiagnostic performance evaluation\u003c/h3\u003e\n\u003cp\u003eDiagnostic performance was assessed across five diseases analyzed as binary classifications and Pfirrmann grading evaluated as a multi-level ordinal classification. SpineNetv2 demonstrated superior overall performance compared with the junior orthopedic surgeon in most pathologies.\u003c/p\u003e\u003cp\u003eSpineNetv2 achieved significantly higher diagnostic accuracy than the junior orthopedic surgeon for CCS (p\u0026thinsp;=\u0026thinsp;0.001), spondylolisthesis (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and bilateral FS (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Performance for herniation was comparable between the two approaches (p\u0026thinsp;=\u0026thinsp;0.293). SpineNetv2 consistently achieved higher specificity and positive predictive values, whereas the junior orthopedic surgeon demonstrated greater sensitivity across most conditions. For Pfirrmann grading, SpineNetv2 showed superior diagnostic concordance, with a significantly lower mean absolute error (0.213 vs. 0.254, p\u0026thinsp;=\u0026thinsp;0.001) compared with the junior orthopedic surgeon (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eError pattern characterization\u003c/h2\u003e\u003cp\u003eSystematic evaluation of SpineNetv2 error patterns across six spinal pathologies showed overall diagnostic accuracy ranging from 83.5\u0026ndash;97.5%, with a mean accuracy of 92.8%. For binary classifications, error directionality was consistent, with false negatives substantially exceeding false positives: CCS (5.4% vs. 0.4%), herniation (5.0% vs. 1.5%), and bilateral FS (4.7\u0026ndash;5.9% vs. 0.2\u0026ndash;0.5%), reflecting the model\u0026rsquo;s conservative diagnostic profile (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e). For Pfirrmann grading, the most common misclassification involved Grade 2 cases predicted as Grade 1 (106 cases, representing 19.3% of all Grade 2 discs). Mild errors (\u0026plusmn;\u0026thinsp;1 grade) accounted for 11.6% of cases, whereas severe errors (\u0026thinsp;\u0026ge;\u0026thinsp;\u0026plusmn;\u0026thinsp;2 grades) occurred in only 4.9% (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eRisk factors for diagnostic discordance\u003c/h2\u003e\u003cp\u003eComprehensive risk stratification across 36 demographic\u0026ndash;spinal combinations (six diseases \u0026times; three age groups \u0026times; two spinal levels) showed predominantly reliable performance of SpineNetv2. Overall diagnostic consistency averaged 92.7% across all subgroups, with risk distribution as follows: 30 low-risk (83.3%), five medium-risk (13.9%), and one high-risk (2.8%) combinations.\u003c/p\u003e\u003cp\u003eRisk concentration was disease-specific, with disc degeneration accounting for all medium- and high-risk scenarios. The only high-risk combination was observed in older patients (\u0026gt;\u0026thinsp;60 years) with upper lumbar disc degeneration (77.7% consistency, 94 cases). The five medium-risk combinations corresponded to the remaining Pfirrmann age\u0026ndash;segment strata (81.6\u0026ndash;86.9% consistency). All other diseases demonstrated consistently high reliability across demographic and segmental subgroups, ranging from 90.8% for CCS to 97.9% for spondylolisthesis (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eMultivariable performance determinants analysis\u003c/h2\u003e\u003cp\u003eFollowing disease consolidation, multivariable analysis was conducted using disc-level logistic regression for binary diseases (12,275 observations) and ordinal logistic regression for Pfirrmann grading (2,455 observations), both with patient-clustered robust standard errors. Significant associations were observed in 5 of 24 factor\u0026ndash;disease combinations (20.8%), with overall accuracy averaging 94.9% across subgroups.\u003c/p\u003e\u003cp\u003eAge effects were limited to Pfirrmann grading, with reduced accuracy in middle-aged (OR\u0026thinsp;=\u0026thinsp;0.54, 95% CI: 0.35\u0026ndash;0.84, p\u0026thinsp;=\u0026thinsp;0.007) and older patients (OR\u0026thinsp;=\u0026thinsp;0.27, 95% CI: 0.15\u0026ndash;0.46, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Male sex was associated with improved accuracy for CCS (OR\u0026thinsp;=\u0026thinsp;1.61, p\u0026thinsp;=\u0026thinsp;0.006) and herniation (OR\u0026thinsp;=\u0026thinsp;1.45, p\u0026thinsp;=\u0026thinsp;0.024). Lower lumbar segments outperformed upper segments only in CCS (OR\u0026thinsp;=\u0026thinsp;1.50, p\u0026thinsp;=\u0026thinsp;0.020). No other factor\u0026ndash;disease combinations demonstrated significant associations (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eIn this independent external validation study, we evaluated the performance of SpineNetv2 in 491 patients, encompassing 2,455 lumbar discs across six common degenerative pathologies. The model achieved high diagnostic accuracy, ranging from 83.5\u0026ndash;97.5% (mean 92.8%), and consistently outperformed the junior orthopedic surgeon in most binary tasks. Accuracy was significantly higher for CCS, spondylolisthesis, and bilateral FS, whereas performance for disc herniation was comparable. For Pfirrmann grading, SpineNetv2 showed superior concordance with a significantly lower MAE (0.213 vs. 0.254), although accuracy declined in older patients and at upper lumbar segments. Taken together, these findings demonstrate a diagnostic profile characterized by high specificity and positive predictive values but relatively lower sensitivity, reflecting a conservative algorithmic tendency that prioritizes minimizing false positives.\u003c/p\u003e\u003cp\u003eEarly deep learning models primarily focused on single diagnostic tasks, with weighted kappa values for Pfirrmann grading ranging from 0.59 to 0.87 [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Recent external validations of SpineNet reported class-balanced accuracies of 74\u0026ndash;79% and kappa values of 0.63\u0026ndash;0.77[\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The higher accuracy observed in our study (average 92.8%) may be explained by several factors. First, we consolidated rare disease grades, thereby simplifying complex multi-class problems into binary tasks. While this likely improved apparent performance, it also reduced granularity in severity assessment. Second, we used overall accuracy as the primary metric rather than class-balanced accuracy, which may overestimate performance under imbalanced disease prevalence. Third, SpineNetv2 may incorporate algorithmic optimizations beyond the original SpineNet. Most importantly, its conservative diagnostic profile, characterized by high specificity and low sensitivity, reduces false positives but risks missing clinically relevant cases. Thus, despite favorable statistical metrics, it is essential to consider how these methodological factors influence the real-world diagnostic utility of the system.\u003c/p\u003e\u003cp\u003eDetailed error analysis revealed a consistent conservative diagnostic pattern across all binary classifications, with false negatives substantially outnumbering false positives. This pattern reflects deliberate specificity prioritization rather than a limitation in capability and is likely driven by several mechanisms. First, class imbalance in the training data, combined with standard decision thresholds, naturally suppresses false positives. Second, the inherent ambiguity of borderline classifications, particularly at transition zones such as Pfirrmann Grades 1\u0026ndash;2 or CCS Grades 1\u0026ndash;2, predisposes the algorithm to underdiagnose uncertain cases. This explains why the most frequent misclassification involved Pfirrmann Grade 2 cases predicted as Grade 1, with most errors limited to \u0026plusmn;\u0026thinsp;1 grade. Third, common imaging confounders, including motion artifacts, endplate sclerosis, Modic changes, and anatomical variations such as short pedicles in the upper lumbar spine, can obscure subtle pathological features, further contributing to conservative predictions[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis diagnostic profile contrasts with the tendency of junior clinicians to over-diagnose, accounting for SpineNetv2\u0026rsquo;s superior positive predictive values despite lower sensitivity. From a clinical perspective, these findings suggest that deployment strategies should prioritize threshold recalibration to improve sensitivity without materially reducing specificity. In practice, the system appears best suited as a confirmatory second reader, particularly valuable for ruling in pathology, whereas negative predictions should not be used as the sole basis for reassurance in high-risk patients or complex cases[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eRisk stratification indicated that nearly all medium- to high-risk scenarios were confined to the Pfirrmann task, while other endpoints remained stable across demographic and segmental strata. The only high-risk subgroup, elderly patients with upper lumbar discs, showed 77.7% accuracy, highlighting the intrinsic limitations of automated ordinal degeneration assessment. In advanced age, composite degenerative changes compress the T2 dynamic range and blur ordinal boundaries between Grades 3\u0026ndash;5. At upper lumbar levels, thinner discs and anatomical crowding increase partial-volume effects and reduce conspicuity. Reader uncertainty is correspondingly higher in these near-boundary states, which is propagated to the model.\u003c/p\u003e\u003cp\u003eTechnical factors further exacerbate this vulnerability. Pfirrmann grading relies on subtle intensity differences, while disc-level context and segment-wise domain shifts can weaken positive evidence. Representation of this subgroup in our cohort was limited, which may hinder calibration and threshold optimization, regardless of the undisclosed training distribution. In contrast, binary endpoints maintained high accuracy across strata, suggesting that presence/absence decisions are less susceptible to these confounders. These findings support the need for age- and segment-conditioned calibration, ordinal-aware training objectives with enriched sampling of elderly upper-lumbar cases, and mandatory expert review for this predefined high-risk group.\u003c/p\u003e\u003cp\u003eThis study has several limitations that should be considered in interpretation. The single-center, single-protocol cohort and reliance on a single-expert reference standard, despite high intra-rater agreement (κ\u0026thinsp;=\u0026thinsp;0.89\u0026ndash;0.94), limit generalizability compared with multi-reader, multi-site designs. Consolidation of rare severe grades stabilized estimation but reduced ordinal granularity and likely inflated apparent accuracy, a notable constraint for pre-surgical stratification where differentiating severe from very severe disease guides treatment decisions. Class imbalance and cohort-specific prevalence further complicate cross-study comparisons, particularly against balanced metrics or alternative endpoints. The absence of clinical correlation restricts conclusions to imaging-based accuracy rather than downstream clinical impact. Sparse data in key strata, most notably elderly upper-lumbar cases, produced imprecise estimates and warrant cautious interpretation, underscoring the need for prospective multicenter studies with multi-reader standards, ordinal endpoints, and sensitivity-oriented calibration.\u003c/p\u003e\u003cp\u003eFuture validation should prioritize multicenter studies with consensus reference standards to confirm SpineNetv2\u0026rsquo;s specificity-oriented diagnostic profile across diverse populations and imaging protocols[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. In the near term, technical refinements should focus on age- and segment-conditioned calibration for Pfirrmann grading and on assessing whether modest threshold relaxation can improve sensitivity without materially reducing specificity, particularly in the high-risk elderly upper-lumbar subgroup[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Beyond algorithmic adjustments, prospective evaluations are needed to determine whether operating at a high-specificity threshold provides tangible clinical benefits, including fewer unnecessary interventions, greater workflow efficiency, and improved diagnostic confidence, when implemented as a second reader. Pending such evidence, our findings support a limited but important role for SpineNetv2 as a confirmatory tool for positive findings rather than a stand-alone diagnostic system[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eSpineNetv2 demonstrated high accuracy across five binary lumbar pathologies, while Pfirrmann grading remained the main limitation, particularly in elderly upper lumbar discs. Its specificity-oriented profile supports use as a confirmatory second reader, but reliance on negative findings is not recommended. Broader reliability will require multicenter, multi-reader validation and sensitivity-oriented calibration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval\u0026nbsp;\u003c/strong\u003eThis study has been approved by the Ethics Committee of our hospital (IRB number KYLL-2025-0012). All procedures involving human participants are conducted in accordance with institutional guidelines, the Helsinki Declaration, and relevant national regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to participate\u0026nbsp;\u003c/strong\u003eAll patients provided written informed consent for the use of their clinical and imaging data for research purposes. Consent was obtained at the time of hospital admission or outpatient consultation using standard institutional consent forms approved by the Ethics Committee. Patients who declined to provide consent were excluded from the study.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eConflicts of interest\u003c/strong\u003e All authors declare that they have no conflict of interest.\u0026nbsp;\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eAvailability of data and material\u0026nbsp;\u003c/strong\u003eThe MRI image data used in this study were obtained from patients at the Second Affiliated Hospital of Zunyi Medical University. Due to patient privacy and institutional regulations, the raw data cannot be made publicly available. Data may be accessed upon reasonable request and with approval from the Institutional Review Board of the Second Affiliated Hospital of Zunyi Medical University.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eAcknowledgements\u003c/strong\u003e I express my sincere gratitude for the unwavering support from my family members.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eAuthor contribution\u0026nbsp;\u003c/strong\u003eConceptualization: Kebing Jin, Qian Du. Data curation: Xingkai Wu, Qianbo Song. Formal Analysis: Jiaxiang Zhou, Zhiyu Zhou, Guangru Cao. Funding acquisition: Zhiyu Zhou, Kebing Jin, Qian Du. Investigation/ Supervision: Zhiyu Zhou, Guangru Cao, Xingkai Wu, Qianbo Song, Jiaxiang Zhou, Kebing Jin, Qian Du. Methodology: Xingkai Wu, Jiaxiang Zhou, Kebing Jin. Project administration: Guangru Cao, Ke bing Jin, Qian Du. Resources: Guangru Cao, Qian Du. Writing - original draft: Xingkai Wu. Writing - review \u0026amp; editing: Qianbo Song, Jiaxaing Zhou, Zhiyu Zhou, Guangru Cao, Kebing Jin, Qian Du. Xingkai Wu and Qianbo Song contributed equally to this work.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eFunding\u0026nbsp;\u003c/strong\u003eThis research was funded by the Basic Research Program of Guizhou Provincial Department of Science and Technology (Qiankehe Foundation-ZK [2024] General-347), the Science and Technology Plan Projects of Zunyi City (Zunshi Kehe HZ (2024) No. 432 and 442), and Guizhou Science and Technology Plan Project (No. QKHRC-KJZY[2025]036).\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eHoy D, March L, Brooks P, Blyth F, et al. (2014) The global burden of low back pain: estimates from the Global Burden of Disease 2010 study.Ann Rheum Dis 73(6):968-974. https://doi.org/10.1136/annrheumdis-2013-204428\u003c/li\u003e\n\u003cli\u003ede Souza I, Sakaguchi TF, Yuan S, Matsutani LA, et al. (2019) Prevalence of low back pain in the elderly population: a systematic review.Clinics (Sao Paulo) 74(e789. https://doi.org/10.6061/clinics/2019/e789\u003c/li\u003e\n\u003cli\u003eKnezevic NN, Candido KD, Vlaeyen J, Van Zundert J, Cohen SP (2021) Low back pain.Lancet 398(10294):78-92. https://doi.org/10.1016/S0140-6736(21)00733-9\u003c/li\u003e\n\u003cli\u003eBeaudart C, McCloskey E, Bruy\u0026egrave;re O, Cesari M, et al. (2016) Sarcopenia in daily practice: assessment and management.BMC Geriatr 16(1):170. https://doi.org/10.1186/s12877-016-0349-4\u003c/li\u003e\n\u003cli\u003eLiang YW, Fang YT, Lin TC, Yang CR, et al. (2024) The Quantitative Evaluation of Automatic Segmentation in Lumbar Magnetic Resonance Images.Neurospine 21(2):665-675. https://doi.org/10.14245/ns.2448060.030\u003c/li\u003e\n\u003cli\u003eXie J, Yang Y, Jiang Z, Zhang K, et al. (2023) MRI radiomics-based decision support tool for a personalized classification of cervical disc degeneration: a two-center study.Front Physiol 14(1281506. https://doi.org/10.3389/fphys.2023.1281506\u003c/li\u003e\n\u003cli\u003eLiawrungrueang W, Park JB, Cholamjiak W, Sarasombath P, Riew KD (2025) Artificial Intelligence-Assisted MRI Diagnosis in Lumbar Degenerative Disc Disease: A Systematic Review.Global Spine J 15(2):1405-1418. https://doi.org/10.1177/ 21925682241274372\u003c/li\u003e\n\u003cli\u003eZheng HD, Sun YL, Kong DW, Yin MC, et al. (2022) Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI.Nat Commun 13(1):841. https://doi.org/10.1038/s41467-022-28387-5\u003c/li\u003e\n\u003cli\u003eChen X, Wang X, Zhang K, Fung KM, et al. (2022) Recent advances and clinical applications of deep learning in medical image analysis.Med Image Anal 79(102444. https://doi.org/10.1016/j.media.2022.102444\u003c/li\u003e\n\u003cli\u003eTsai JY, Hung IY, Guo YL, Jan YK, et al. (2021) Lumbar Disc Herniation Automatic Detection in Magnetic Resonance Imaging Based on Deep Learning.Front Bioeng Biotechnol 9(708137. https://doi.org/10.3389/fbioe.2021.708137\u003c/li\u003e\n\u003cli\u003eBeulah A, Sharmila TS, Pramod VK (2021) Degenerative disc disease diagnosis from lumbar MR images using hybrid features.The Visual Computer 38(8): https://doi.org/10. 1007/ s00371-021-02154-x\u003c/li\u003e\n\u003cli\u003eLehnen NC, Haase R, Faber J, R\u0026uuml;ber T, et al. (2021) Detection of Degenerative Changes on MR Images of the Lumbar Spine with a Convolutional Neural Network: A Feasibility Study.Diagnostics (Basel) 11(5):902. https://doi.org/10.3390/diagnostics11050902\u003c/li\u003e\n\u003cli\u003eWang ZX, Hu YG (2012) High-intensity zone (HIZ) of lumbar intervertebral disc on T2-weighted magnetic resonance images: spatial distribution, and correlation of distribution with low back pain (LBP).Eur Spine J 21(7):1311-1315. https://doi.org/10.1007/ s00586-012-2240-0\u003c/li\u003e\n\u003cli\u003eYi W, Zhao J, Tang W, Yin H, et al. (2023) Deep learning-based high-accuracy detection for lumbar and cervical degenerative disease on T2-weighted MR images.Eur Spine J 32(11):3807-3814. https://doi.org/10.1007/s00586-023-07641-4\u003c/li\u003e\n\u003cli\u003eAbdollah V, Parent EC, Batti\u0026eacute; MC (2019) Reliability and validity of lumbar disc height quantification methods using magnetic resonance images.Biomed Tech (Berl) 64(1):111-117. https://doi.org/10.1515/bmt-2017-0086\u003c/li\u003e\n\u003cli\u003eWindsor, R., Jamaludin, A., Kadir, T, et al. (2022) SpineNetV2: Automated Detection, Labelling and Radiological Grading Of Clinical MR Scans. ArXiv, abs/2205.01683. https://doi.org/10.48550/arXiv.2205.01683\u003c/li\u003e\n\u003cli\u003eWindsor R, Jamaludin A, Kadir T, Zisserman A (2024) Automated detection, labelling and radiological grading of clinical spinal MRIs.Sci Rep 14(1):14993. https://doi.org/10.1038/s41598-024-64580-w\u003c/li\u003e\n\u003cli\u003eBurke JF, Sussman JB, Kent DM, Hayward RA (2015) Three simple rules to ensure reasonably credible subgroup analyses.BMJ 351(h5651. https://doi.org/10.1136/bmj.h5651\u003c/li\u003e\n\u003cli\u003eChen XL, Li XY, Wang Y, Lu SB (2023) Relation of lumbar intervertebral disc height and severity of disc degeneration based on Pfirrmann scores.Heliyon 9(10):e20764. https://doi.org/10.1016/j.heliyon.2023.e20764\u003c/li\u003e\n\u003cli\u003eAndreisek G, Imhof M, Wertli M, Winklhofer S, et al. (2013) A systematic review of semiquantitative and qualitative radiologic criteria for the diagnosis of lumbar spinal stenosis.AJR Am J Roentgenol 201(5):W735-746. https://doi.org/10.2214/AJR.12.10163\u003c/li\u003e\n\u003cli\u003eKoslosky E, Gendelberg D (2020) Classification in Brief: The Meyerding Classification System of Spondylolisthesis.Clin Orthop Relat Res 478(5):1125-1130. https://doi.org/10.1097/CORR.0000000000001153\u003c/li\u003e\n\u003cli\u003eBaur D, Bieck R, Berger J, Sch\u0026ouml;fer P, et al. (2025) Automated Three-Dimensional Imaging and Pfirrmann Classification of Intervertebral Disc Using a Graphical Neural Network in Sagittal Magnetic Resonance Imaging of the Lumbar Spine.J Imaging Inform Med 38(2):979-987. https://doi.org/10.1007/s10278-024-01251-2\u003c/li\u003e\n\u003cli\u003eGrob A, Loibl M, Jamaludin A, Winklhofer S, et al. (2022) External validation of the deep learning system \u0026quot;SpineNet\u0026quot; for grading radiological features of degeneration on MRIs of the lumbar spine.Eur Spine J 31(8):2137-2148. https://doi.org/10.1007/s00586-022-07311-x\u003c/li\u003e\n\u003cli\u003eMcSweeney TP, Tiulpin A, Saarakkala S, Niinim\u0026auml;ki J, et al. (2023) External Validation of SpineNet, an Open-Source Deep Learning Model for Grading Lumbar Disk Degeneration MRI Features, Using the Northern Finland Birth Cohort 1966.Spine (Phila Pa 1976) 48(7):484-491. https://doi.org/10.1097/BRS.0000000000004572\u003c/li\u003e\n\u003cli\u003eNigru AS, Benini S, Bonetti M, Bragaglio G, et al. (2024) External validation of SpineNetV2 on a comprehensive set of radiological features for grading lumbosacral disc pathologies.N Am Spine Soc J 20(100564. https://doi.org/10.1016/j.xnsj.2024.100564\u003c/li\u003e\n\u003cli\u003eNikpasand M, Middendorf JM, Ella VA, Jones KE, et al. (2024) Automated magnetic resonance imaging-based grading of the lumbar intervertebral disc and facet joints.JOR Spine 7(3):e1353. https://doi.org/10.1002/jsp2.1353\u003c/li\u003e\n\u003cli\u003eZhao P, Zhu S (2025) Advances and challenges in AI-assisted MRI for lumbar disc degeneration detection and classification.Eur Spine J. https://doi.org/10.1007/s00586-025-09179-z\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-spine-journal","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"esjo","sideBox":"Learn more about [European Spine Journal](http://link.springer.com/journal/586)","snPcode":"586","submissionUrl":"https://submission.springernature.com/new-submission/586/3","title":"European Spine Journal","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Lumbar spine, Magnetic resonance imaging, Degenerative pathology, SpineNetv2, Artificial intelligence, Diagnostic performance, External validation, Pfirrmann grading","lastPublishedDoi":"10.21203/rs.3.rs-7559680/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7559680/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eMagnetic resonance imaging (MRI) is the reference standard for evaluating degenerative lumbar spine disorders, but interpretation is time-consuming and subject to inter-observer variability. SpineNetv2, a publicly available deep learning system, enables automated analysis of multiple spinal pathologies. This study conducted an independent external validation of SpineNetv2 against expert reference standards.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eA total of 491 patients (2,455 lumbar discs, L1/2\u0026ndash;L5/S1) were retrospectively included. Disc-level reference labels were established by an expert orthopedic surgeon, with a junior orthopedic surgeon serving as comparator. Six pathologies were assessed: disc degeneration (Pfirrmann grading), central canal stenosis (CCS), spondylolisthesis, herniation, and bilateral foraminal stenosis (FS). Performance metrics included sensitivity, specificity, positive predictive value, negative predictive value, F1-score, Matthews correlation coefficient, exact agreement, weighted kappa, and mean absolute error. McNemar\u0026rsquo;s test and bootstrap resampling (1,000 iterations) were used for statistical analysis.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eOverall accuracy ranged from 83.5\u0026ndash;97.5% (mean 92.8%). SpineNetv2 significantly outperformed the junior orthopedic surgeon in CCS, spondylolisthesis, and bilateral FS (all p\u0026thinsp;\u0026le;\u0026thinsp;0.001), with comparable performance in herniation (p\u0026thinsp;=\u0026thinsp;0.293). Pfirrmann grading showed lower MAE for SpineNetv2 compared with the junior surgeon (0.213 vs. 0.254, p\u0026thinsp;=\u0026thinsp;0.001), though accuracy declined in older patients and upper lumbar discs. Error analysis revealed a specificity-oriented profile, with false negatives exceeding false positives.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eSpineNetv2 demonstrated high accuracy across five binary lumbar pathologies, while Pfirrmann grading remained the main limitation, particularly in elderly upper lumbar discs. Its specificity-oriented profile supports use as a confirmatory second reader, but reliance on negative findings is not recommended. Broader reliability will require multicenter, multi-reader validation and sensitivity-oriented calibration.\u003c/p\u003e","manuscriptTitle":"External Validation of SpineNetv2 Deep Learning System for Automated Lumbar Spine MRI Analysis: A Multi-pathology Diagnostic Accuracy Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-22 08:53:17","doi":"10.21203/rs.3.rs-7559680/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-08T16:11:26+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-06T14:14:49+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-25T20:40:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"150762686591110911747940924797358008251","date":"2025-09-25T15:11:21+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"32016797146500198841387361673430269242","date":"2025-09-23T13:57:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"258556253585609673028795637008675265728","date":"2025-09-10T16:47:56+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-10T15:30:39+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-09T05:37:15+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-09T05:34:57+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Spine Journal","date":"2025-09-08T03:38:56+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-spine-journal","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"esjo","sideBox":"Learn more about [European Spine Journal](http://link.springer.com/journal/586)","snPcode":"586","submissionUrl":"https://submission.springernature.com/new-submission/586/3","title":"European Spine Journal","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"2f028bdd-f5e8-41be-9225-e018912ef921","owner":[],"postedDate":"September 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-11-10T16:08:13+00:00","versionOfRecord":{"articleIdentity":"rs-7559680","link":"https://doi.org/10.1007/s00586-025-09543-z","journal":{"identity":"european-spine-journal","isVorOnly":false,"title":"European Spine Journal"},"publishedOn":"2025-11-08 15:57:42","publishedOnDateReadable":"November 8th, 2025"},"versionCreatedAt":"2025-09-22 08:53:17","video":"","vorDoi":"10.1007/s00586-025-09543-z","vorDoiUrl":"https://doi.org/10.1007/s00586-025-09543-z","workflowStages":[]},"version":"v1","identity":"rs-7559680","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7559680","identity":"rs-7559680","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}