Deep learning for automated spinopelvic parameter measurement from radiographs: a meta-analysis

INTRODUCTION

Quantitative evaluation of spine and pelvis anatomy has long interested clinicians and researchers in fields such as orthopedics, neurosurgery, and radiology. Assessing sagittal spinal balance - the geometric relationships between spinal curves and pelvic parameters - is considered essential for understanding normal posture and alignment^[1]. Sagittal balance encompasses important radiographic measures such as cervical and lumbar lordosis, thoracic kyphosis, pelvic tilt (PT), pelvic incidence (PI), and sacral slope (SS)^[2,3]. Abnormal spinopelvic alignment has been associated with pain, disability, and poor health outcomes^[4].

Traditionally, spinopelvic parameters were manually measured from plain radiographs using techniques like the Cobb method, with known limitations in accuracy and objectivity^[5]. Computer-assisted analysis tools later emerged to potentially improve measurement consistency, though substantial human input was still required^[6]. Deep learning has rapidly advanced in recent years but traces its origins back decades. The concepts of neural networks were initially developed in the 1950s and 60s. However, computational power limited applications. In the 1980s and 90s, techniques like convolutional neural networks (CNNs) were pioneered, laying the groundwork for modern deep learning. Major advancements in computing, along with the availability of large datasets, then enabled deep neural networks to surpass previous benchmarks across diverse tasks. Beginning in the 2010s, deep learning achieved remarkable performance in computer vision, natural language processing, and medical imaging analysis. The latest methods like CNNs now offer transformative opportunities to extract information from complex data. Over the past decade, advances in artificial intelligence and machine learning have enabled more automated approaches for quantitative radiology and medical imaging^[7,8].

Machine learning utilizes statistical models trained on known data to recognize patterns in new data^[9]. Deep learning is a subset of machine learning based on layered neural networks that can automatically learn optimal features directly from raw data, unlike traditional techniques requiring hand-crafted feature engineering^[10]. The latest deep learning methods have become integral for the automated analysis of medical images across specialties^[11,12], including quantitative characterization of spine disorders from radiographs and CT scans^[13,14].

Several recent studies have applied deep CNNs for automated measurement of key spinopelvic parameters from standard radiographs^[15]. Reported accuracy has been promising but varies widely across studies. However, a comprehensive synthesis of the latest achievements, methodological innovations, and measured performance has been lacking. This review aims to systematically summarize and critically appraise the existing literature on deep learning-based assessment of sagittal spinopelvic alignment on radiographs. It elucidates the current state of the field and future directions to potentially improve clinical adoption.

METHODS

This Meta-analysis was conducted according to the preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines^[16] [Supplementary Table 1].

Study selection

Studies were included if they met the following criteria: (1) published in English language peer-reviewed journals; (2) used deep learning models including CNNs to automatically estimate spinopelvic parameters from radiographs (X-ray); (3) reported model performance metrics compared to human rater measurements including mean absolute error and correlation coefficient. Conference abstracts, case reports, editorials, and non-peer reviewed articles were excluded.

Two reviewers (A.K.M and J.C) independently screened the titles, abstracts, and full texts of retrieved records against the eligibility criteria. Disagreements were resolved by consensus or consultation with a third reviewer if needed. The study selection process was documented using a PRISMA flow diagram [Figure 1].

Figure 1. PRISMA flow chart. PRISMA: Preferred reporting items for systematic reviews and meta-analyses.

RESULTS

A total of 14 studies published between 2018-2023 were included in this systematic review, encompassing 10,727 subjects^[18-31]. The studies utilized various imaging modalities to develop and validate automated methods for measuring spinal alignment, including lateral X-rays^[18-31], biplanar radiographs^[30,31], and CT scans^[26]. Both preoperative and postoperative images were employed, with 6 studies incorporating cases with spinal implants^{[19,20,23,24,28]} to evaluate performance in surgically altered anatomy. The diversity of imaging captures numerous clinically relevant scenarios, although multicenter external validation was lacking, with most datasets from single institutions [Table 1].

A range of deep learning models were applied for automated spinal measurement, including custom CNNs^{[18-21,23-25,27-31]}, multi-view correlation networks^[19,20], and segmentation-based approaches^{[23,24,28-30]}. For Cobb angle measurement, mean absolute errors ranged from 1.2° to 7.81°^{[18-21,23,27,28]}, with most studies achieving errors ≤ 5°. Similar trends were observed for other sagittal parameters, such as thoracic kyphosis, lumbar lordosis, and PI^{[18,19,22-25,29,30]}. Intraclass correlation coefficients between automated and manual measurements exceeded 0.75, indicating strong agreement^[22,25,26]. Computational efficiency was reported in several studies, with inference times ranging from 0.2 to 75 s per image^{[22,23,27,28]}, demonstrating the potential for accelerated analysis compared to manual methods.

Cobb angle demonstrated a pooled mean error of 4.3° (95%CI: 3.2°-5.4°). Thoracic kyphosis and lumbar lordosis showed similar pooled errors of 3.9° (95%CI: 2.7°-5.1°) and 3.6° (95%CI: 2.8°-4.4°), respectively. PT had the lowest pooled error at 1.9° (95%CI: 1.3°-2.5°), while PI exhibited a slightly higher pooled error of 4.1° (95%CI: 2.7°-5.5°). Sagittal vertical axis (SVA) demonstrated a pooled mean error of 1.3 cm (95%CI: 0.9-1.7 cm). These results highlight the overall accuracy of deep learning models in automatically measuring key spinopelvic parameters from radiographic images [Figure 2].

Figure 2. Forest plot showing weight distribution of the different spino-pelvic parameters.

Manual measurement of spinopelvic parameters has shown inter-observer variability ranging from 5° to 10° for Cobb angle measurements and similar ranges for other parameters in previous studies. The pooled AI measurement errors we found (4.3° for Cobb angle, 3.9° for thoracic kyphosis, and 3.6° for lumbar lordosis) demonstrate comparable or better accuracy than manual measurements while offering significantly improved efficiency.

DISCUSSION

This systematic review and meta-analysis demonstrate the potential of deep learning for the automated measurement of spinopelvic parameters from radiographs. The comprehensive literature search identified 14 eligible studies between 2018-2023, analyzing over 10,000 radiographs with deep CNNs and other architectures^[18-31]. The studies utilized various imaging sources to develop and validate automated methods for measuring spinal alignment, including lateral X-rays^{[19,20,23-25,27,32-34]}, biplanar radiographs^[31,35], and CT scans^[36]. Both preoperative^{[19,23,25,27,31-33,35,36]} and postoperative^[19,20,24] images were employed, with 6 studies incorporating cases with spinal implants^{[19,20,23,24,33,36]} to evaluate performance in surgically altered anatomy. The diversity of imaging captures numerous clinically relevant scenarios. However, multicenter external validation was lacking, with most datasets from single institutions. Aspects like vendor variability could impact segmentation. Model development must be capable of analyzing all imaging protocols for translation.

A range of model types were applied for automated spinal measurement, from conventional machine learning^[21,25] and rule-based systems^[31] to modern deep CNNs^{[19,20,23-25,31,33-37]}. Details for replication varied extensively - 4 studies provided no specific model details^{[19,25,27,34]}, while 5 gave networks and parameters^{[19,20,25,31,35]}. Public code/data availability remains limited. Custom architectures were common for direct spinal measurement^{[19,33-35,37]}, rather than off-the-shelf models. Multi-task^[25,33,37], multi-view^[19,33], and vertebral correlation^[25] learning schemes showed benefits for parameter accuracy through inter-relationship modeling, overcoming imaging challenges like occlusion.

Studies assessed accuracy via comparison to expert manual measurement, using metrics such as mean absolute differences (all studies) and voxel overlap measures where segmentation was evaluated^{[19,23,24,31,35,36]}. For Cobb angle measurement, mean errors ranged from 1.7° to 8.1°, but most CNN methods achieved ≤ 5° mean difference^{[23-25,31,32,34,35]}, adequate for clinical usage^[38]. Similar trends were held for other sagittal measurements^{[19,20,23,24,31,35]}. Notably, Wang et al. employed extrapolation methods atop initial estimates to give the best overall accuracies of 6.2°/7.8° Cobb angle errors in lateral/AP views vs. 4.0°/4.1° for MCV-Net^[20,37]. Intraclass coefficients of 0.86-0.99^[19,23-25] confirmed automated/manual measurement agreement.

Comparisons were made to traditional manual measurement^{[19,20,23-25,31,35]}, manual tools^[19,25,27], early machine learning applications^[25], and different iterations of automated algorithms^[19,33]. Automated methods met or exceeded both classic and contemporary techniques. Particular benefits arose in reproducibility, efficiency, and standardization vs. manual approaches prone to subjectivity and variability^[19,23,24]. Deep learning methods showed headroom over alternate automated implementations in accuracy, overcoming limitations such as occlusion. Wang et al. achieved better Cobb measurement than MCV-Net^[19] (7.8° lateral error vs. 4.1°), through vertebral correlation and extrapolation augmentations^[20].

Studies cited small datasets^[31], external validity^{[19,24,31,35,36]}, surgical cases^{[19,20,23,24,33]}, implant handling^[33,36], need for inter-rater evaluations^[33], pelvic measurement gaps^[27], follow-up studies^[24], and real-world clinical workflow integration^[24,27] as main limitations. Anonymization, reproducibility, negative societal impacts, and public data availability were generally not addressed. Small samples particularly restricted subgroup analysis - only Gami et al. reported metrics by spinal pathology^[27]. Building large heterogeneous benchmark datasets could facilitate model development and address generalizability. Standardized reporting guidelines for spine AI could also benefit the field.

End-to-end runtimes ranged from 2 to 75 s for automated measurement pipelines^[23,24], up to 17× faster than manual analysis; most systems took under 20 s^[19,23,35], adequate for surgical usage. Inference-only times were often sub-second^[23,27]. Accelerated measurement enables more intraoperative images for improved surgical decisions. However, detailed computational profiling was generally lacking, impeding comparisons. Cloud-based implementations could broadly enable these techniques.

Studies used statistical comparisons between automated and manual measurements for validation, incorporating Bland-Altman analysis^{[19,23,25,27,31,35]}, paired significance tests^{[19,23,27,35]}, linear regression^{[19,23,25,27,31,35]}, Pearson correlation coefficients^{[19,23,25,27,31,35]}, and intra-class coefficients^[19,23-25]. Manual measurement reliability was sometimes quantified^[27]. Both preoperative^{[19,23-25,27,31,32,35]} and postoperative subjects^[19,24,37] were included, although only Kim et al. performed validation in distinct pre- and postoperative cohorts^[24]. Most evaluations used held-out testing data from the same institution as model development; multicenter validation was absent. Generalizability beyond the typically homogeneous training populations requires further scrutiny.

CNN backbones ranged from VGG^[19] and U-Nets^[31,36] to ResNets^[24,25,33]. Both feedforward^[19,25] and fully convolutional layouts were used. Custom network engineering was common^{[19,23-25,27,31,32,35]}, given insufficient anatomical representational power in generic classification architectures. Pretraining on natural images via Mask R-CNN^[36] and DetectNet^[34] helped offset smaller target dataset sizes. Segmentation-based approaches employed secondary algorithms on CNN outputs to estimate spinal parameters^{[24,25,31,35,36]}, adding measurement variability. End-to-end sagittal measurement could minimize error propagation within integrated networks.

Reported batch sizes during neural network training spanned 16-256. However, 10 studies did not specify this optimization detail at all^{[18-31,34,36]}. Small batches can enhance generalization and reduce overfitting, but at a computational cost. Larger batches offer efficiency yet may miss anomalous cases. Standardization would benefit reproducibility. The median batch size was 64^{[24,31,33,36]}, aligning with typical practices.

The number of training epochs ranged from 30 to 6,000 for deep neural networks. But again, most studies omitted specifics^{[18-31,34,36]}. Two reports described adaptive epoch counts based on validation improvements^[24,36], rather than fixed values. Typical regimes were 30-50 epochs^[31,33]. Standardized detail would aid reproducibility. Generalizability with shorter training requires scrutiny where transfer learning was not employed.

The IJMEDI checklist for medical imaging AI highlighted several shortcomings (see tabulated results in requests), particularly around enabling reproducibility. Areas such as software details, computational resource usage, model accessibility, and evaluation set specificity suffered poor reporting. However, studies did well in conveying overall aims, statistical and evaluation methodology, and limitations. Recent initiatives for standardizing ML reporting^[39,40], plus reproducibility checklists^[38], may benefit new spine AI imaging research.

Despite promising accuracy, certain limitations remain. Most studies used single-institutional data lacking sufficient diversity^{[19-21,23,25,28-30]}. Reference standards from manual radiograph measurements intrinsically incorporate subjectivity from inter-observer variations^[41]. CT imaging remains unevaluated. Studies for some parameters are still few. Real-world clinical validation is lacking^[42]. Our subgroup analyses found that studies using CNN architectures demonstrated higher accuracy for parameters like lumbar lordosis compared to other models. This highlights the importance of selecting appropriate architectures tailored to the specific radiographic quantification task. As deep learning continues advancing, further research is still needed to optimize model design and determine the most effective architectures for automated spinopelvic measurement. Larger comparative studies evaluating different network architectures on common datasets would help elucidate the relative merits and guide selection.

Moving forward, larger multicenter studies should validate these models before clinical implementation^[40,43]. Continued research on handling label noise and measurement uncertainty is required^[13,41]. Standardized reporting guidelines could enhance reproducibility^[40]. Models should be optimized across diverse settings and pathologies^[42,43]. Clinically meaningful accuracy metrics deserve focus beyond errors^[41].

The application of deep learning models and their potential role in spine surgery has already begun to be explored. Of value to spine surgeons, models have demonstrated success in diagnosing various musculoskeletal and spinal disorders, including sarcopenia, scoliosis, and low back pain^[44-47]. In regard to prognosis, deep learning models have been successful in predicting postoperative complications such as surgical site infections and 30-day readmission rates after lumbar fusion procedures^[48,49]. While these initial findings are promising, further research validating the use of these models in other realms of patient care, particularly surgical planning, is needed.

Spinopelvic parameters are of great importance to the surgeon for planning, and methods of measurement have evolved significantly. Early assessments began with the Cobb angle and focused on spinal curvatures but overlooked the pelvis. In the 1980s and 1990s, the introduction of parameters such as PI, PT, and SS revolutionized the understanding of sagittal balance. These measurements linked pelvic alignment to spinal posture. By the 2000s, global spinal alignment gained attention, with the SVA and newer measures like the T1 pelvic angle (TPA) becoming essential for surgical planning in adult spinal deformity (ASD).

Up until the early 2000s, measurement of spinopelvic parameters was mostly done manually and, on average, took 3-15 min. The manual measurement process is tedious and time-consuming while also being prone to rater-dependent error^[50]. Advancements in imaging techniques, including full-body electron optic system (EOS) radiographs, CT scans, and MRI, have enabled more accurate measurements of spinopelvic parameters. The development of more sophisticated software has led to accelerated measurement times via semi-automated computer-aided tools, such as SurgiMap^[50]. Software tools such as SurgiMap have demonstrated a mean time efficiency of 75 ± 25 s to perform a full spinopelvic analysis, significantly reducing the burden associated with manual measurements^[50]. Our review of the existing literature on deep learning models for spinopelvic parameter measurement revealed processing times ranging from 0.2 to 1 s per image. A set of radiographs for spinopelvic parameter measurement typically involves 2-3 images on average: a lateral X-ray, an anterior and posterior X-ray, and possibly a full-body EOS image in more complicated cases. Regarding time saved, deep learning models would require an estimated 0.6-3 s to analyze a full set of images compared to the 75-second mean from the studies mentioned previously. Deep learning models are, therefore, roughly 25× more efficient. Additionally, there were studies included in our analysis that involved pathological images, whereas the study using SurgiMap involved images with no pathology, further demonstrating the capability and efficiency of deep learning technology. To contextualize these efficiency gains with accuracy: Manual measurements typically show inter-observer variability of 5°-10° for the Cobb angle and similar ranges for other parameters. Semi-automated tools reduced this variability to 3°-7°. Our meta-analysis found AI measurement errors of 4.3° for Cobb angle, 3.9° for thoracic kyphosis, and 3.6° for lumbar lordosis - comparable to or better than both manual and semi-automated methods. This suggests AI can dramatically improve measurement efficiency without compromising accuracy, potentially offering both time savings and measurement reliability improvements in clinical practice.

No one model stood out as superior to the others. Each study and the model they used had advantages and disadvantages that are open to interpretation. For example, the model used by Zerouali et al. was mainly tested in a pediatric population; therefore, this model would likely only be of interest to a surgeon who operates on this population^[22]. Many studies only involved a single clinical dataset, which is a key reason why we argue for multicenter validation to demonstrate reproducibility. Additionally, some studies did not train their models on patients who had implants. Therefore, these models would require further validation to be useable in scenarios such as postoperative evaluation and planning for revision surgery. What was consistent across all models was that they all were more efficient than current methods without compromising accuracy.

Despite the demonstrated accuracy and efficiency of these models, there remains a gap in understanding their practical utility for surgeons across various clinical contexts, including preoperative and intraoperative stages. Theoretically, the enhancement in efficiency should offer surgeons more time to review images and make surgical plans. Pending multicenter validation, future research should explore whether or not the integration of deep learning truly enhances efficiency throughout the entire perioperative continuum. For example, a surgeon may use deep learning as an adjunct for formulating a preoperative plan. Within surgery, intraoperative X-ray image evaluation may allow synchronous measurement of spinopelvic parameters to assess the efficacy of hardware placement. Lastly, in the postoperative phase, the technology can be used to predict postoperative complications and 30-day readmission rates as stated earlier, with the potential for much more. No one model stood out as superior to the others. Each study and the model it used had advantages and disadvantages that are open to interpretation.

A notable limitation in measuring PI deserves specific attention. Our meta-analysis found PI measurements had a relatively higher pooled error of 4.1° compared to other pelvic parameters such as PT (1.9°). This larger error can be attributed to several specific challenges: First, the presence of double-dome endplates can make it difficult to precisely identify the sacral endplate angle. Second, femoral head overlapping, particularly in patients with high BMI or osteoarthritis, can obscure the precise center of the femoral head. Third, the quality of lateral radiographs, especially in patients with wide pelvises, can result in poor visualization of anatomical landmarks due to increased soft tissue density. Fourth, metallic implants such as total hip replacements can create artifacts that interfere with landmark identification. These factors compound measurement uncertainty and likely contribute to the higher error rates observed for PI across studies. Future deep learning models should specifically address these challenges, perhaps through specialized preprocessing steps or architectural modifications designed to better handle landmark obscurity and anatomical variations.

As this technology continues to evolve, it is highly unlikely that it will not play a role in patient healthcare. It is of great importance for future research to ensure adequate ethical standards, as new concepts and technologies are often met with some resistance. Issues with accountability, transparency, and permissions could come into question by involving deep learning in the decision-making process. Therefore, the integration of deep learning technology should come as a complementary tool in the surgical decision-making processes, where surgeons can potentially optimize patient care pathways and improve overall clinical outcomes.