Accurate automatic measurement of spinopelvic parameters with a one-stage deep learning technique

Accurate automatic measurement of spinopelvic parameters with a one-stage deep learning technique | 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 Accurate automatic measurement of spinopelvic parameters with a one-stage deep learning technique Xianglong Meng, Jianhua Liu, zihe feng, Yu Sun, Zhiheng Zhao, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3734310/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: The current method of measuring parameters in spinal imaging manually is time-consuming and prone to inconsistencies. This study proposed and validated a novel method to automate the measurement of pelvic parameters using a one-stage deep learning (DL) model. Methods: Spinopelvic parameters, including pelvic incidence (PI), sacral slope (SS), and pelvic tilt (PT), were measured from full body radiographs of patients by three evaluators and by using our proposed method. Our proposed one-stage DL model was based on keypoint localisation. Landmark localisation error was used to evaluate the performance of landmark localisation. To evaluate the agreement between our method and the human evaluators, the analysis of average error, standard deviation, and intra- and inter-evaluator reliability was conducted using the intraclass correlation coefficient (ICC) and Pearson's correlation coefficient ( R ). Results: The method achieved excellent measurement performance for spinopelvic parameters. The distribution of the landmark localisation errors was within a reasonable range (median error, 2.28–4.01 mm). ICC values for the assessment of the intra- (range: 0.941–0.996) and inter-evaluator (0.994–0.998) reliability of human evaluators were excellent. The method was able to determine spinopelvic parameters with excellent ICC values (0.919-0.997) and R value ( R >0.899, p <0.001, all). Meanwhile, the detection speed of the algorithm was approximately 30 times faster than that of manual measurements of spinopelvic parameters. Conclusions: This one-step automated measurement method is less time-consuming and has excellent reliability and agreement with human evaluators. spinopelvic parameters deep learning landmark localisation parameter measurement Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Background Sagittal balance in patients is a critical metric in the diagnosis, treatment, surgical planning, and subsequent analysis of spinal disease[ 1 , 2 ]. Sacral slope (SS), pelvic tilt (PT), and pelvic incidence (PI) are the most commonly measured spinopelvic sagittal balance parameters. They are closely related to pain, function, and health-related quality of life outcomes[ 3 – 6 ]. However, two significant challenges exist in measuring pelvic parameters in clinical practice, as follows: 1) novice physicians lack measurement experience and are prone to making errors; 2) the process is time-consuming[ 7 – 9 ]. Artificial intelligence (AI) has shown potential in providing simple, fast, and accurate measurement methods[ 10 , 11 ]. Recently, many studies have utilised deep learning (DL) techniques to measure sagittal parameters. However, there are still some limitations to these methods. First, training with a relatively small dataset can lead to model overfitting and poor generalisation ability, which cannot effectively assist doctors in making correct diagnoses[ 12 , 13 ]. Second, the measurement process is divided into two stages—object detection and keypoint detection—which, although highly accurate, is time-consuming[ 14 , 15 ]. Finally, the evaluation metrics of the model performance need to be more comprehensive to effectively assess the difference between the AI measurement results and surgeon annotations[ 16 – 18 ]. This study aimed to develop and validate a novel method to fully automate the measurement of pelvic parameters, including PI, PT, and SS. To this end, this study used a one-stage model to improve measurement accuracy and generalisation capabilities to assist doctors in making relevant diagnoses. Methods This study was approved by the ethics committee of China Beijing Chaoyang hospital, Capital Medical Univercity and the requirement for informed consent was waived due to its retrospective nature. the approval document number is 2021-8-3-1. Experimental materials We collected imaging data from 1,193 patients with full-length lateral spine X-ray images taken using the GE Definium 8000 system (General Electric Company, Boston, Massachusetts, MA, USA) at the Beijing Chaoyang Hospital between 2018 and 2021. We eliminated 193 unqualified radiographs from the total image data based on the integrity of the femoral head, and the remaining 1,000 images comprised an effective dataset. The basic information of the dataset included 631 and 369 radiography images from male and female patients, respectively. The mean age of the patients was 56 (range, 35–85) years. Proposed methods The key activities performed in constructing a new automatic spinopelvic parameter measurement system are illustrated in Fig. 1 . The valid dataset was preprocessed by cropping and resizing and subsequently split into training and testing sets in a ratio of 8:2 (i.e., 800 and 200 radiographs, respectively). The training and testing sets were used as inputs for the automatic measurement model to complete model training and testing. The model uses keypoint detection technology to determine the relative positions of the spine and femoral heads and calculate spinopelvic parameters (PT, PI, SS). Three orthopaedic surgeons manually annotated each radiograph in the testing set, identifying keypoint (anterior edge, centre of the S1 endplates, and centres of the femoral heads) coordinates and measuring pelvic parameters, which were used as reference standards for subsequent experiments. The landmark location model was subsequently trained to obtain the optimal model, and its performance was subsequently tested. The landmark location error was evaluated to compare the keypoint of the model with expert annotations. The accuracy of the model's spinopelvic parameter measurements was analysed by comparing them with the experts’ measurements. The automatic measurement system of spinopelvic parameters performed the following steps to measure spinopelvic parameters on a radiographic image, as shown in Fig. 2 . First, the position of keypoints was located through landmark localisation. Subsequently, spinopelvic parameters were located based on the relative positions of keypoints. During the aforementioned steps, landmark localisation requires selecting a suitable object detection model to locate the positions of the spine and femoral heads quickly and accurately. Parameter measurements require adding auxiliary lines according to the positions of keypoints and using geometric theory to solve the value of the spinopelvic parameter. The details are elaborated as follows. Radiography dataset preprocessing It is necessary to preprocess the X-ray image to meet the input requirements of the landmark location model, as depicted in Fig. 3 (a). Preprocessing comprised two parts: cropping and resizing. The bottom half (1824×1932) was derived from the full X-ray image and resized to 512×512. Data labelling is another important task. All X-ray images were manually labelled for key landmarks to generate the ground truth for model training. A detailed illustration of the location of the four key landmarks on a radiograph and corresponding information are shown in Fig. 3 (b). The two keypoints—S1L and S1C—represent the leading edge and centre of the S1 endplate. The other two keypoints—FH1 and FH2—define the centre of the femoral head. The data labelling task was completed by two surgeons using LabelMe software version 5.1.1 (MIT, Boston, Massachusetts, MA, USA) and confirmed by an expert, which guaranteed the accuracy of keypoint labelling. Landmark localisation model Compared with the existing research, the landmark location model based on CenterNet[ 19 ]. In this paper a one-stage anchor-free object detection method is used that achieves a trade-off between accuracy and speed. The architecture of the Landmark Location model comprises two parts: the backbone and the output heads (Fig. 4 ). Using DLA-34[ 20 ] as the network backbone for the model, four output heads predicted the centre heatmap, centre offset, corner heatmap, and corner offset. The result of the two offset heads was used to compensate for the coordinate deviation caused by down-sampling rounding. The trained model could automatically infer the positions of the four keypoints in a radiograph that had not been marked with keypoints (Fig. 4 ). The loss function included heatmap prediction loss and offset regression loss. The former, an improved version of focal loss, measured the disparity between the predicted heatmap and the actual target location. The latter utilised L1 norm loss to measure the difference between the predicted and actual offsets. Meanwhile, data augmentation technology was implemented during model training to improve the model's accuracy and reduce data overfitting. Variations in brightness, contrast, and sharpness may be present in radiographs obtained from different machines, which can impact the precision and generalisability of the model. Therefore, data augmentation techniques were employed to effectively tackle this problem. We applied various modifications (such as rotated, flipped, and cropped) to the radiograph to simulate the variations in angle, direction, and exposure time during image acquisition. Specifically, we rotated, flipped, and cropped the images. In addition, we adjusted the contrast and brightness of the images to simulate differences in exposure time and added noise to simulate the variations in the acquisition process. Parameter measurement Spinopelvic parameters can be calculated from the four predicted key landmarks by adding appropriate auxiliary lines. We fitted two lines and made four additional lines based on the four key landmarks, as shown in Fig. 2 (c): one was the connecting line between the centres of the two femoral heads (whose midpoint was called the hip axis [HA]), and the other was along the upper edge of the sacral (S1) endplate; the four auxiliary lines included a horizontal reference, a vertical reference, a line connecting the HA with the centre of the S1 endplate, and a line orthogonal to the S1 endplate. Finally, we measured the three spinopelvic parameters PI, PT, and SS using geometric relationships among the six lines. Model training The dataset comprised the training (800 images) and test (200 images) sets. The former was used to train the model using the Pytorch framework[ 21 ] on an NVIDIA GeForce GTX 3090Ti GPU for fitting the parameters of the model. The latter was used to evaluate the performance of the model by comparing its results with manually measured results. The model was initialised with publicly available pre-trained weights, the learning rate was set to 0.001, and the optimiser chose Adam. During training, the model was trained for 100 epochs. Statistical analysis Our method was evaluated using the test dataset (200 images) for landmark localisation error and spinopelvic parameter estimation. Localisation error refers to the difference between the predicted position of a keypoint and the actual position annotated by the expert. The test set was fed into the landmark localisation model to predict the position coordinates of four keypoints. Since the local errors were non-normally distributed, we used a boxen plot to visualise error distributions. To evaluate the intra- and inter-evaluator reliability of manual and model-based measurements of spinopelvic parameters, we enlisted two resident physicians (Evaluators 1 and 2) and one attending surgeon (Evaluator 3) to measure the spinopelvic parameters in the test dataset, with Evaluators 1 and 2 performing one measurement each, and Evaluator 3 conducting two measurements (Evaluator 3a and 3b), which were used as the gold standard for comparison. The angles of spinopelvic parameters in the test set were measured using the SurgiMap Spine software (Nemaris, Methuen, Massachusetts, MA, USA). All measurements were performed independently and blindly, meaning that after Evaluator 3 completed the first measurement, a second measurement was performed two weeks later. Moreover, the results of one evaluator were not eligible for a review by another evaluator. These manually measured values were used as the ground truth to compare with those obtained using our method (AI method) predictions and hence validate the algorithm. For spinopelvic parameter estimation, we employed four evaluation strategies: mean absolute difference (MAD) and its corresponding standard deviation (SD), the mean of errors between the measurement and the gold standard (ME), intraclass correlation coefficient (ICC), and Pearson correlation coefficient (R)[ 22 – 24 ]. ICCs were used to evaluate the agreement between the model- and expert-measured values. ICCs between 0.75 and 1.00 were interpreted as having excellent reliability (good, 0.60–0.74; fair, 0.40–0.59; and poor, < 0.40)[ 25 ]. Pearson correlation coefficients were used to evaluate the correlations between the predicted spinopelvic parameters and their corresponding ground truth values. Bland–Altman (B&A) plots were used to estimate the consistency intervals between the expert- and model-measured results[ 26 ]. In this study, all statistical analyses were conducted using Python version 3.3.6. Results Performance of landmark localisation The localisation errors of the keypoints are visualised using the boxen plots in Fig. 5 . The localisation error of the centre of the S1 endplate (S1C) was the smallest (median error, 1.74 mm), and the median localisation error for the anterior edge of the S1 endplate (S1L) was approximately 2.03 mm, as shown in Fig. 5 (a). When abnormal samples were removed from the test set, we observed that the landmark location error of the anterior edge on S1L was the smallest among the four landmarks (median error, 2.28 mm), as shown in Fig. 5 (b). Parameter measurement result The intra- and inter-evaluator reliability analyses of manual measurements are shown in Table 1 . ICCs for the inter-evaluator reliability were always the smallest for SS (0.945, Evaluator 1 vs Evaluator 3a) and the largest for PT (0.996, Evaluator 2 vs Evaluator 3a). R values were the largest for PT (0.995, Evaluator 2 vs Evaluator 3a); MAD values were also the largest for PT (2.143, Evaluator 2 vs Evaluator3a). The ME values were smaller than 2.76°, with the largest values for SS (Evaluator 1 vs Evaluator 3a). For the inter-evaluator reliability, ICCs were the smallest for SS and largest for PT. Additionally, all MAD values were less than 0.65. Table 1 Intra- and inter-evaluator reliability of the manual measurements for spinopelvic parameters Statistical method PI PT SS Inter-evaluator reliability Evaluator 1 vs Evaluator 3a MAD (SD) 1.591 (1.104) 0.829 (0.582) 1.571 (1.039) ME 2.63 ± 1.94 1.56 ± 1.01 2.56 ± 1.88 R ( p -value) 0.943 (< \({10}^{-87}\) ) 0.985 (< \({10}^{-138}\) ) 0.922 (< \({10}^{-75}\) ) ICC 0.965 (< \({10}^{-92}\) ) 0.995 (< \({10}^{-168}\) ) 0.955 (< \({10}^{-83}\) ) Evaluator 1 vs Evaluator 3b MAD (SD) 1.273 (1.075) 0.500 (0.429) 1.313 (1.088) ME 2.19 ± 1.67 1.46 ± 0.66 2.12 ± 1.70 R ( p -value) 0.959 (< \({10}^{-85}\) ) 0.996 (< \({10}^{-172}\) ) 0.949 (< \({10}^{-78}\) ) ICC 0.964 (< \({10}^{-91}\) ) 0.994 (< \({10}^{-160}\) ) 0.954(< \({10}^{-}\) ) Evaluator 2 vs Evaluator 3a MAD (SD) 1.582 (1.290) 0.425 (0.295) 1.646(1.210) ME 2.62 ± 2.04 0.76 ± 0.52 2.76 ± 2.04 R ( p -value) 0.943 (< \({10}^{-86}\) ) 0.995 (< \({10}^{-179}\) ) 0.934 (< \({10}^{-80}\) ) ICC 0.960 (< \({10}^{-86}\) ) 0.996 (< \({10}^{-180}\) ) 0.952 (< \({10}^{-79}\) ) Evaluator 2 vs Evaluator 3b MAD (SD) 1.423 (1.254) 0.379 (0.301) 1.401 (1.132) ME 2.39 ± 1.90 0.47 ± 0.38 2.46 ± 1.80 R ( p -value) 0.941 (< \({10}^{-85}\) ) 0.994 (< \({10}^{-171}\) ) 0.930 (< \({10}^{-78}\) ) ICC 0.959 (< \({10}^{-85}\) ) 0.996 (< \({10}^{-172}\) ) 0.949 (< \({10}^{-78}\) ) Evaluator 1 vs Evaluator 2 MAD (SD) 1.871 (1.346) 0.859 (0.591) 1.853 (1.372) R ( p -value) 0.933 (< \({10}^{-80}\) ) 0.984 (< \({10}^{-134}\) ) 0.920 (< \({10}^{-73}\) ) ICC 0.951 (< \({10}^{-79}\) ) 0.989 (< \({10}^{-135}\) ) 0.941 (< \({10}^{-72}\) ) Intra-evaluator reliability Evaluator 3a vs Evaluator 3b MAD (SD) 0.450 (0.409) 0.374 (0.305) 0.611 (0.607) R ( p -value) 0.996 (< \({10}^{-190}\) ) 0.997 (< \({10}^{-200}\) ) 0.991 (< \({10}^{-160}\) ) ICC 0.997 (< \({10}^{-192}\) ) 0.998 (< \({10}^{-201}\) ) 0.994 (< \({10}^{-162}\) ) MAD, mean absolute difference; ME, mean error; SD, standard deviation; ICC, intraclass correlation coefficient; PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope For the inter-evaluator reliability, analysis between the AI model and manual measurements was conducted, as shown in Table 2 . ICC values for spinopelvic parameters ranged between 0.919 and 0.997, with the smallest values for SS (AI method vs Evaluator 2) and the largest for PT (AI method vs Evaluator 3a). The MAD values were the smallest for PT (0.367, Evaluator 3a vs AI) and the largest for SS (2.097, Evaluator 2 vs AI method). The R values for spinopelvic parameters were the smallest for SS (0.89, evaluator vs AI). Table 2 Inter-evaluator reliability of the manual measurements and AI method for spinopelvic parameters Statistical method PI PT SS AI method vs Evaluator 1 MAD (SD) 1.318 (1.180) 0.978(0.613) 1.354 (1.167) ME 2.22 ± 1.77 2.18 ± 1.15 1.98 ± 1.79 R ( p -value) 0.948 (< \({10}^{-90}\) ) 0.982 (< \({10}^{-131}\) ) 0.924 (< \({10}^{-76}\) ) ICC 0.964 (< \({10}^{-91}\) ) 0.987 (< \({10}^{-132}\) ) 0.924 (< \({10}^{-77}\) ) AI method vs Evaluator 2 MAD (SD) 1.931 (1.263) 0.437 (0.275) 2.097 (1.324) ME 3.40 ± 2.31 0.75 ± 0.52 3.97 ± 2.48 R ( p -value) 0.931 (< \({10}^{-78}\) ) 0.992 (< \({10}^{-162}\) ) 0.899 (< \({10}^{-72}\) ) ICC 0.949 (< \({10}^{-77}\) ) 0.994 (< \({10}^{-163}\) ) 0.919 (< \({10}^{-67}\) ) AI method vs Evaluator 3a MAD (SD) 0.941 (0.772) 0.367 (0.299) 1.035 (0.753) ME 1.47 ± 1.22 0.62 ± 0.47 1.74 ± 1.28 R ( p -value) 0.976 (< \({10}^{-120}\) ) 0.995 (< \({10}^{-187}\) ) 0.965 (< \({10}^{-106}\) ) ICC 0.983 (< \({10}^{-120}\) ) 0.997 (< \({10}^{-188}\) ) 0.975 (< \({10}^{-105}\) ) AI method vs Evaluator 3b MAD (SD) 0.975 (0.793) 0.376 (0.299) 1.059 (0.814) ME 1.50 ± 1.26 0.61 ± 0.48 1.86 ± 1.33 R ( p -value) 0.975 (< \({10}^{-119}\) ) 0.995 (< \({10}^{-186}\) ) 0.961 (< \({10}^{-101}\) ) ICC 0.983 (< \({10}^{-119}\) ) 0.997 (< \({10}^{-187}\) ) 0.972 (< \({10}^{-100}\) ) PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope; AI, artificial intelligence Compared to manual measurements, the AI method showed the smallest errors for PI, TT, and SS when compared to Evaluator 3a (gold standard), with values of 0.16, 0.27, and 1.13, respectively, as shown in Table 3 . Exemplarily, the correlation and agreement between AI and Evaluator 3a for spinopelvic parameters are further depicted in Fig. 6 . Figure 7 presents the detection success rate of the parameters, which refers to the proportion of test cases with the measured MAD lower than the threshold. Specifically, the AI-measured results were close to the mean of the expert manual measurement results while exhibiting a smaller variance. The evaluator in this study reported an average time of 2 to 3 minutes for measuring spinopelvic parameters per image, whereas the proposed algorithm accomplished the same analysis in approximately 5 seconds. Table 3 Angle measurement by the manual measurements and AI method for spinopelvic parameters PI PT SS AI Method 45.76°±8.00° (29.60°, 60.96°) 10.03°±6.41° (-2.52°, 22.58°) 35.26°±6.61° (22.58°, 48.44°) Evaluator 1 44.05°±7.81° (28.26°, 61.57°) 7.88°±6.45° (-2.53°, 22.90°) 34.62°±6.71° (21.48°, 47.76°) Evaluator 2 47.75°±8.58° (30.92°, 64.57°) 10.44°±6.34° (-1.98°, 22.85°) 38.71°±7.57° (23.86°, 53.54°) Evaluator 3a 45.92°±8.33° (29.59°, 62.24°) 9.52°±6.37° (-2.97°, 22.00°) 36.39°±7.02° (22.62°, 50.15°) Evaluator 3b 46.03°±8.27° (27.63°, 60.20°) 9.52°±6.38° (-2.97°, 22.02°) 36.52°±7.02° (22.76°,50.26°) PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope Discussion Sagittal balance plays a vital role in the diagnosis, treatment, and surgical planning of spinal disorders. We presented an automated method to measure pelvic parameters and compared it with expert manual measurements. Excellent agreement was demonstrated between human measurement and the automated method for the spinopelvic parameters (PI, PT, and SS). It can be observed that the localisation errors of F1 and F2 are larger than those of the anterior edge and the centre of S1 in Fig. 5 (b). As an X-ray is taken from the lateral spine, there was an inevitable overlap of the femoral head. The distant femoral head is easily occluded, resulting in the false marking of keypoints. The smaller the landmark error, the closer the position of the keypoint is to the real position, and the more accurate the calculated parameter value. Overall, the localisation error of the four key landmarks was < 8.0 mm for most X-rays in the test set, an acceptable error range for expert surgeon measurements. The measurement of parameters in spinal imaging is time-consuming and particularly prone to inconsistency. Therefore, several studies have evaluated intraobserver and interobserver variability in sagittal spinopelvic balance parameters measured manually from radiographs with computer-aided tools[ 3 ]. Malioot et al. reported an SD of 3.5° when measuring SS, PT, and PI using the Keops software (SMAIO, Lyon, France)[ 8 ]. Lafage et al. evaluated the efficacy of the SurgiMap Spine software (Nemaris, Methuen, Massachusetts, MA, USA) and reported intraobserver standard deviations of 4.6°, 2.5°, 0.8°, and 5.4° for SS, PT, and an angle similar to ST and PI, respectively[ 9 ]. Vila-Casademunt et al. also evaluated the SurgiMap Spine software by measuring intraobserver MAD and interobserver MAD for spinopelvic balance parameters. They observed that the presence of lumbosacral instrumentation resulted in a significant increase in interobserver variability[ 27 ]. Although parameter measurement errors can be reduced using computer-aided tools, they are still classified as manual measurements owing to the use of the mouse to mark points and draw lines and other geometric shapes, including circles, on the X-ray. The measurement of pelvic parameters continued to be affected by the X-ray quality and observer experience, and the validity and reliability of parameter measurement could not be guaranteed. The emergence of AI has made automated measurement tools more intelligent and automated, which not only helps avoid repetitive manual measurement but also saves the time cost in clinical practice. Galbusera et al. proposed a DL model based on the location of keypoints for spinopelvic parameter prediction[ 14 ], which calculated the difference between AI and the manual measurements of PT, PI, and SS, yielding MADs (SD) of is 2.7° (0.7°), 9.5° (8.5°), and 6.9° (6.2°), respectively. Korez et al. proposed a two-stage model based on the U-Net network structure to implement a DL measurement tool for the fully automated measurement of spinopelvic parameters[ 17 ]. The MADs (SD) between manual measurements and automated tools for PT, PI, and SS were 2.7° (2.5°), 1.2° (1.2°), and 5.0° (3.4°), respectively. Furthermore, decentralised convolutional DL-based approaches have been proven to significantly improve parameter measurement accuracy. Chae et al. presented an automated method for measuring spinopelvic parameters with reported MADs of 1.45°, 2.51°, and 2.64° for PT, PI, and SS, respectively[ 28 ]. The method proposed by Nguyen et al. had MADs (SD) of 1.16° (0.82°), 2.21° (2.05°), and 3.17° (3.09°) for PT, PI, and SS, respectively, representing the lowest deviations compared with other methods reported in the previous literature[ 29 ]. Our method enables the fully automatic measurement of spinopelvic parameters using a simple one-stage DL model. Compared to other methods, this approach achieves higher efficiency by directly performing one-stage detection through keypoint detection. It shortens the measurement time through a single training and eliminates the need for multi-stage detection. More importantly, our method achieves better inter-class agreement with manual measurements. To validate the measurement quality of human evaluators in evaluating the AI algorithm, intra- and inter-evaluator measurements were performed. PT showed the highest reliability among the spinopelvic parameters between human evaluators (Tables 1 ), while SS had relatively lower ICCs values. The results indicate high reliability within the three evaluators (ICC > 0.93), which is well in line with previously published assessments[ 30 ] and demonstrates the validity of manual measurements as the ground truth to evaluate AI algorithms. It proved the reliability of measuring parameters on radiographic images for the one-stage DL model. The agreement between the AI method and one attending surgeon's measurements was higher than that between one attending surgeon and two resident physicians' measurements. For PT, PI, and SS, the ICCs and the R value between AI and Evaluator 3a, as well as Evaluator 3b, were higher when compared to the corresponding values derived from Evaluator 1 and Evaluator 2. Moreover, the MAD values between AI and Evaluator 3a, as well as Evaluator 3b, were smaller in comparison to the values computed between Evaluator 1 and Evaluator 2, as evident from Table 2 . In the meantime, we analysed the consistency between AI and Evaluator 3a (gold standard). The ICCs ranged between 0.975 and 0.997. R was between 0.965 and 0.995, and MAD (SD) was 0.941° (0.772°) for PI, 0.367° (0.299°) for PT, and 1.035° (0.753°) for SS in Table 2 . Exemplarily, the MAD (SD) for AI vs Evaluator 3a was lower than the values for instance reported by Nguyen et al.[ 29 ] for the spinopelvic parameters (PI 0.941°<2.21°; PT 0.367°<1.16°; SS 1.035°<3.17°). Moreover, the correlation scatter diagrams in Fig. 6 (a), (c), and (e) and the B&A diagrams in Fig. 6 (b), (d), and (f) further showed that the results measured using our method were in good agreement with the results marked by the attending surgeon. Compared to the other two parameters, the consistency of AI and expert measurement in measuring SS was slightly lower, but still within an acceptable margin of error for experts. As demonstrated in the B&A diagrams, the differences in measurement results followed a normal distribution; 95% of the differences should fall within the MD ± 1.96 SD interval, known as the 95% limits of agreement (95% LoA), with only a few points out of bounds. This indicates that the AI-based measurement method is highly consistent with the results obtained by human evaluator measurements. Our method had a higher measurement accuracy and shorter time overhead. Compared to the measurements obtained by two resident physicians, the measurement accuracy of the AI method was closer to that obtained by one attending surgeon on two separate occasions in Table 3 . The measurement time for all spinopelvic parameters in each case using this method has been reduced by approximately 30 times compared to manual measurements, resulting in a significant improvement in the detection speed. Furthermore, we analysed the detection success rate, which means that when the error angle was smaller than the threshold, the detection value was considered acceptable. In Fig. 7 , when the degree threshold was set to 3°, the detection success rate exceeded 80%. Experts can recognise and accept this threshold angle. Of note, this threshold angle fell within an acceptable margin of error for experts and did not have any impact on the final diagnostic results. Our approach demonstrated higher reliability compared to other methods, as evidenced by the statistical analysis conducted using manual measurements from multiple experts. When compared to expert manual measurements, our method exhibited superior detection accuracy. The evaluation of measurement time validated the higher measurement speed of our method. Moreover, we collected a larger number of X-ray data samples as the training and testing sets. Increasing the number of X-ray data samples for the training dataset could improve the AI model in extracting rich feature information to accurately identify keypoints and calculate the pelvic parameters. Additionally, a larger number of test data samples could help ensure the fairness of the model evaluation process and verify the generalisation capabilities of the AI model. However, this study has some limitations. First, experts still exhibited certain biases in making judgements based on small amounts of complex data. Second, the model could not effectively correct the errors caused by states such as coronal deformities and anatomical differences, resulting in measurement errors. In the future, we intend to leverage knowledge transfer techniques from the field of AI to acquire feature knowledge from complex samples, thereby reducing measurement errors. Conclusions Our study proposed a new method for the fully automatic measurement of PI, PT, and SS spinopelvic parameters via the DL model-driven identification of key landmarks. The statistical analyses were built based on the deviation of the proposed model predictions versus measurements by three evaluators who conducted manual measurement using a test dataset of 1,000 radiographs. Our method has been proven to possess high measurement accuracy, excellent consistency, and faster measurement speed for spinopelvic parameter measurement. It effectively alleviated laborious routine tasks, saving valuable time and reducing the likelihood of error-prone manual measurements. Furthermore, its potential applicability extends to other orthopaedic sub-domains. Abbreviations AI Artificial intelligence DL Deep learning HA Hip axis ICCV International Conference on Computer Vision MD Mean difference SD Standard deviation ICC Intraclass correlation coefficient MAD Mean absolute difference PI Pelvic incidence PT Pelvic tilt Declarations Ethics approval and consent to participate: This study was approved by the ethics committee of China Beijing Chaoyang hospital, Capital Medical Univercity and the requirement for informed consent was waived due to its retrospective nature. The approval document number is 2021-8-3-1. Consent for publication Not applicable. Availability of data and materials The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Competing interests The authors (Xianglong Meng, Jianhua Liu, Zihe Feng , Yu Sun, Zhiheng Zhao, Zhiqiang Bai, and Yong Hai) declare no conflict of interest. Funding This work was supported by the R&D Program of Beijing Municipal Education Commission (NO. KZ202210025034), the Science and Technology and Information Technology Bureau of Chaoyang District, Beijing (CN) (NO. CYSF2019), and the National Natural Science Foundation of China ((NO. 6227070318). Authors’ contributions * Xianglong Meng, Jianhua Liu, and Zihe Feng 1 contributed equally to this study. Conceptualization: Xianglong Meng, Yu Sun; Methodology: Jianhua Liu, Zihe Feng ; Formal analysis and investigation: Jianhua Liu, Zihe Feng, Zhiheng Zhao; Writing-original draft preparation: Jianhua Liu, Zihe Feng ; Writing-review and editing: Xianglong Meng, Yu Sun, Jianhua Liu, Zihe Feng ; Supervision: Xianglong Meng, Yu Sun, Zhiqiang Bai, Yong Hai. All authors approved the manuscript. References Azimi P, Yazdanian T, Benzel EC, Hai Y, Montazeri A: Sagittal balance of the cervical spine: A systematic review and meta-analysis . Eur Spine J 2021, 30 :1411-1439. Schlösser TP, Castelein RM, Grobost P, Shah SA, Abelin-Genevois K: Specific sagittal alignment patterns are already present in mild adolescent idiopathic scoliosis . Eur Spine J 2021, 30 :1881-1887. Vrtovec T, Janssen MM, Likar B, Castelein RM, Viergever MA, Pernuš F: A review of methods for evaluating the quantitative parameters of sagittal pelvic alignment . Spine J 2012, 12 (5):433-446. Korez R, Putzier M, Vrtovec T: Automated measurement of pelvic incidence from X-ray images . In . Cham: Springer International Publishing; 2019: 146-152. Korez R, Putzier M, Vrtovec T: Computer-assisted measurement of sagittal pelvic alignment parameters from radiographic images . Zdravniski Vestnik 2018, 87 (11-12):519-529. Garbossa D, Pejrona M, Damilano M, Sansone V, Ducati A, Berjano P: Pelvic parameters and global spine balance for spine degenerative disease: the importance of containing for the well being of content . Eur Spine J 2014, 23 (6):616-627. Krupinski EA: Current perspectives in medical image perception . Atten Percept Psychophys 2010, 72 (5):1205-1217. Maillot C, Ferrero E, Fort D, Heyberger C, Le Huec JC: Reproducibility and repeatability of a new computerized software for sagittal spinopelvic and scoliosis curvature radiologic measurements: Keops® . Eur Spine J 2015, 24 (7):1574-1581. Lafage R, Ferrero E, Henry JK, Challier V, Diebo B, Liabaud B, Lafage V, Schwab F: Validation of a new computer-assisted tool to measure spino-pelvic parameters . Spine J 2015, 15 (12):2493-2502. Charles YP, Lamas V, Ntilikina Y: Artificial intelligence and treatment algorithms in spine surgery . Orthop Traumatol Surg Res 2023, 109 (1, Supplement):103456. Azimi P, Yazdanian T, Benzel EC, Aghaei HN, Azhari S, Sadeghi S, Montazeri A: A review on the use of artificial intelligence in spinal diseases . Asian Spine J 2020, 14 (4):543. Aubert B, Vazquez C, Cresson T, Parent S, Guise JAd: Toward automated 3D spine reconstruction from biplanar radiographs using CNN for statistical spine model fitting . IEEE Trans Med Imaging 2019, 38 (12):2796-2806. Orosz L, Haines CM, Thomson A, Schuler TC, Good CR, Grover P, Dreischarf M, Roy R, Jazini E: 74. Novel artificial intelligence algorithm can accurately and independently measure spinopelvic parameters . Spine J 2021, 21 :S36-S37. Galbusera F, Niemeyer F, Wilke H-J, Bassani T, Casaroli G, Anania C, Costa F, Brayda-Bruno M, Sconfienza LM: Fully automated radiological analysis of spinal disorders and deformities: a deep learning approach . Eur Spine J 2019, 28 (5):951-960. Wu H, Bailey C, Rasoulinejad P, Li S: Automated comprehensive adolescent idiopathic scoliosis assessment using MVC-Net . M Image Anal 2018, 48 :1-11. Weng C-H, Wang C-L, Huang Y-J, Yeh Y-C, Fu C-J, Yeh C-Y, Tsai T-T: Artificial intelligence for automatic measurement of sagittal vertical axis using ResUNet framework . J Clin Med 2019, 8 (11):1826. Korez R, Putzier M, Vrtovec T: A deep learning tool for fully automated measurements of sagittal spinopelvic balance from X-ray images: performance evaluation . Eur Spine J 2020, 29 (9):2295-2305. Cho BH, Kaji D, Cheung ZB, Ye IB, Tang R, Ahn A, Carrillo O, Schwartz JT, Valliani AA, Oermann EK et al : Automated measurement of lumbar lordosis on radiographs using machine learning and computer vision . Global Spine J 2020, 10 (5):611-618. Duan K, Bai S, Xie L, Qi H, Huang Q, Tian Q: CenterNet: Keypoint Triplets for Object Detection . In: 2019 IEEE/CVF International Conference on Computer Vision. IEEE; 2019: 6568-6577. Zhou X, Koltun V, Krähenbühl P: Tracking Objects as Points . In: European Conference on Computer Vision. Cham: Springer; 2020: 474-490. Choi J, Yeom HY, Kim Y: Improving Oversubscribed GPU Memory Performance in the PyTorch Framework . Cluster Computing 2022. Cicchetti DV: Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology . Psychological Assessment 1994, 6 :284-290. WEIR JP: Quantifying test-retest reliability using the intraclass correlation coefficient and the SEM . The Journal of Strength & Conditioning Research 2005, 19 (1):231-240. Cohen I, Huang Y, Chen J, Benesty J, Benesty J, Chen J, Huang Y, Cohen I: Pearson correlation coefficient . Noise reduction in speech processing 2009:1-4. Harrison DE, Cailliet R, Harrison DD, Janik TJ, Holland B: Reliability of Centroid, Cobb, and Harrison Posterior Tangent Methods: Which to Choose for Analysis of Thoracic Kyphosis . Spine 2001, 26 (11):e227-e234. Altman DG, Bland JM: Measurement in medicine: the analysis of method comparison studies . J R Stat Soc Ser A Stat Soc: Series D 1983, 32 (3):307-317. Vila-Casademunt A, Pellisé F, Acaroglu E, Pérez-Grueso FJS, Martín-Buitrago MP, Sanli T, Yakici S, de Frutos AG, Matamalas A, Sánchez-Márquez JM et al : The reliability of sagittal pelvic parameters: the effect of lumbosacral instrumentation and measurement experience . Spine 2015, 40 (4):E253-E258. Chae D-s, Nguyen TP, Park S-J, Kang K-Y, Won C, Yoon J: Decentralized convolutional neural network for evaluating spinal deformity with spinopelvic parameters . Comput Methods Programs Biomed 2020, 197 :105699. Nguyen TP, Jung JW, Yoo YJ, Choi SH, Yoon J: Intelligent evaluation of global spinal alignment by a decentralized convolutional neural network . J Digit Imaging 2022, 35 (2):213-225. Somoskeöy S, Tunyogi-Csapó M, Bogyó C, Illés T: Accuracy and reliability of coronal and sagittal spinal curvature data based on patient-specific three-dimensional models created by the EOS 2D/3D imaging system . Spine J 2012, 12 (11):1052-1059. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted 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-3734310","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":263674223,"identity":"d4e7f65c-24af-4eee-b581-c30e0c102cf6","order_by":0,"name":"Xianglong Meng","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xianglong","middleName":"","lastName":"Meng","suffix":""},{"id":263674224,"identity":"8cb838d0-ae46-4f2e-91f3-ea66a2671ab7","order_by":1,"name":"Jianhua Liu","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Jianhua","middleName":"","lastName":"Liu","suffix":""},{"id":263674226,"identity":"5a23c4b6-31e5-423b-8d73-9382948e3f64","order_by":2,"name":"zihe feng","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"zihe","middleName":"","lastName":"feng","suffix":""},{"id":263674228,"identity":"07d5e1a1-1507-4d70-89f9-8f2ebd0d4949","order_by":3,"name":"Yu Sun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxklEQVRIiWNgGAWjYNCCigMQmod4LWdI1sLYRooW+Rm5x6R5591J7G8/wPjgbRuDvDkhLQY38tKkebc9S5xxJoHZcG4bg+HOBkJaJHLMJGduO5zYcCCBTZq3jSHB4ABBh4G0zDmcOP/8A/bfRGlhuJFjJvGx4XDihhsJbMxEaTE488bY4sOxZ8YbbzxslpxzTsJwA0GHtecY3kiouSM773zywQ9vymzkCTuMgYFFAkIzNgAJCcLqgYD5A1HKRsEoGAWjYOQCACrgREcKq3BhAAAAAElFTkSuQmCC","orcid":"","institution":"Beihang University","correspondingAuthor":true,"prefix":"","firstName":"Yu","middleName":"","lastName":"Sun","suffix":""},{"id":263674229,"identity":"9b40df84-a85e-45d4-91df-6fdd818fd557","order_by":4,"name":"Zhiheng Zhao","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhiheng","middleName":"","lastName":"Zhao","suffix":""},{"id":263674230,"identity":"f1f1ff3c-ae08-4ea0-a327-7ff229f19292","order_by":5,"name":"Zhiqiang Bai","email":"","orcid":"","institution":"Beijing First Hospital of Integrated Chinese and Western Medicine","correspondingAuthor":false,"prefix":"","firstName":"Zhiqiang","middleName":"","lastName":"Bai","suffix":""},{"id":263674231,"identity":"6f4e8b17-b647-4be9-b625-fbde18582643","order_by":6,"name":"Yong Hai","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yong","middleName":"","lastName":"Hai","suffix":""}],"badges":[],"createdAt":"2023-12-10 13:44:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3734310/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3734310/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":49081607,"identity":"1b1fbc59-bab6-49be-82a3-51d57a686b72","added_by":"auto","created_at":"2024-01-02 20:08:11","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":746473,"visible":true,"origin":"","legend":"\u003cp\u003eDesign process of the automatic measurement system for spinopelvic parameters\u003c/p\u003e","description":"","filename":"floatimage1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/b9402753927dc5cd114fb030.jpg"},{"id":49081091,"identity":"067b4010-4c2c-495b-950e-5a73314e0d9f","added_by":"auto","created_at":"2024-01-02 20:00:11","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":217950,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the automatic measurement system for spinopelvic parameters\u003c/p\u003e","description":"","filename":"floatimage2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/2364b315399a605d0dd16ee3.jpg"},{"id":49081093,"identity":"0f76853c-c61b-4be9-80ad-45fedbbb089c","added_by":"auto","created_at":"2024-01-02 20:00:11","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":268246,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the original X-ray preprocessing\u003c/p\u003e","description":"","filename":"floatimage3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/6d6efdf6fae232bc1880ce85.jpg"},{"id":49081608,"identity":"dd42513f-ae12-453d-a16d-c2216fe57cb5","added_by":"auto","created_at":"2024-01-02 20:08:11","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":390564,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the landmark localisation model\u003c/p\u003e","description":"","filename":"floatimage4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/248a23d8065edc8efa1a268e.jpg"},{"id":49081097,"identity":"c32a74e7-dcfe-4e4b-9bbc-fb3ef3bebe59","added_by":"auto","created_at":"2024-01-02 20:00:11","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":176920,"visible":true,"origin":"","legend":"\u003cp\u003eLandmark location error of four landmark coordinates. The value of the median error is marked using boldfont digits within the boxplots. (a) Absolute localisation errors for the whole test set (mm). (b) Landmark location errors for the entire test set with abnormal samples removed (mm).\u003c/p\u003e","description":"","filename":"floatimage5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/b839f11aa92a56832e90602a.jpg"},{"id":49081094,"identity":"a27829b0-c2ec-404c-9ee3-45563f31c21e","added_by":"auto","created_at":"2024-01-02 20:00:11","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":947250,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation scatter diagrams (left column, a, c, and e) for measuring spinopelvic parameters. The Bland–Alman diagrams (right column, b, d, and f) show the discrepancies between the automated measurement system and expert surgeon's standard values when measuring spinopelvic parameters for PT, PI, and SS. The horizontal lines of mean difference (MD) (solid blue line) and MD±1.96 SD (red dashed lines) in Fig. PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope; SD, standard deviation\u003c/p\u003e","description":"","filename":"floatimage6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/245e9963c1ff3907ceaaea76.jpg"},{"id":49081095,"identity":"c92ae3cd-8f65-4904-9e6d-30caf7785997","added_by":"auto","created_at":"2024-01-02 20:00:11","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":267294,"visible":true,"origin":"","legend":"\u003cp\u003eDetection rate of PI, PT, and SS against different degrees of error threshold. PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope\u003c/p\u003e","description":"","filename":"floatimage7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/4eefbb77dbecc9ccf568eb67.jpg"},{"id":69686739,"identity":"95d1ae02-3579-4838-9519-bc2148d4beaa","added_by":"auto","created_at":"2024-11-23 08:46:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4317708,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3734310/v1/4fef529d-8a70-4c5e-8e03-783442919175.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Accurate automatic measurement of spinopelvic parameters with a one-stage deep learning technique","fulltext":[{"header":"Background","content":"\u003cp\u003eSagittal balance in patients is a critical metric in the diagnosis, treatment, surgical planning, and subsequent analysis of spinal disease[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Sacral slope (SS), pelvic tilt (PT), and pelvic incidence (PI) are the most commonly measured spinopelvic sagittal balance parameters. They are closely related to pain, function, and health-related quality of life outcomes[\u003cspan additionalcitationids=\"CR4 CR5\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, two significant challenges exist in measuring pelvic parameters in clinical practice, as follows: 1) novice physicians lack measurement experience and are prone to making errors; 2) the process is time-consuming[\u003cspan additionalcitationids=\"CR8\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI) has shown potential in providing simple, fast, and accurate measurement methods[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Recently, many studies have utilised deep learning (DL) techniques to measure sagittal parameters. However, there are still some limitations to these methods. First, training with a relatively small dataset can lead to model overfitting and poor generalisation ability, which cannot effectively assist doctors in making correct diagnoses[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Second, the measurement process is divided into two stages\u0026mdash;object detection and keypoint detection\u0026mdash;which, although highly accurate, is time-consuming[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Finally, the evaluation metrics of the model performance need to be more comprehensive to effectively assess the difference between the AI measurement results and surgeon annotations[\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study aimed to develop and validate a novel method to fully automate the measurement of pelvic parameters, including PI, PT, and SS. To this end, this study used a one-stage model to improve measurement accuracy and generalisation capabilities to assist doctors in making relevant diagnoses.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis study was approved by the ethics committee of China Beijing Chaoyang hospital, Capital Medical Univercity and the requirement for informed consent was waived due to its retrospective nature. the approval document number is 2021-8-3-1.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eExperimental materials\u003c/h2\u003e \u003cp\u003eWe collected imaging data from 1,193 patients with full-length lateral spine X-ray images taken using the GE Definium 8000 system (General Electric Company, Boston, Massachusetts, MA, USA) at the Beijing Chaoyang Hospital between 2018 and 2021. We eliminated 193 unqualified radiographs from the total image data based on the integrity of the femoral head, and the remaining 1,000 images comprised an effective dataset. The basic information of the dataset included 631 and 369 radiography images from male and female patients, respectively. The mean age of the patients was 56 (range, 35\u0026ndash;85) years.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eProposed methods\u003c/h3\u003e\n\u003cp\u003eThe key activities performed in constructing a new automatic spinopelvic parameter measurement system are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The valid dataset was preprocessed by cropping and resizing and subsequently split into training and testing sets in a ratio of 8:2 (i.e., 800 and 200 radiographs, respectively). The training and testing sets were used as inputs for the automatic measurement model to complete model training and testing. The model uses keypoint detection technology to determine the relative positions of the spine and femoral heads and calculate spinopelvic parameters (PT, PI, SS). Three orthopaedic surgeons manually annotated each radiograph in the testing set, identifying keypoint (anterior edge, centre of the S1 endplates, and centres of the femoral heads) coordinates and measuring pelvic parameters, which were used as reference standards for subsequent experiments. The landmark location model was subsequently trained to obtain the optimal model, and its performance was subsequently tested. The landmark location error was evaluated to compare the keypoint of the model with expert annotations. The accuracy of the model's spinopelvic parameter measurements was analysed by comparing them with the experts\u0026rsquo; measurements.\u003c/p\u003e \u003cp\u003eThe automatic measurement system of spinopelvic parameters performed the following steps to measure spinopelvic parameters on a radiographic image, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e. First, the position of keypoints was located through landmark localisation. Subsequently, spinopelvic parameters were located based on the relative positions of keypoints. During the aforementioned steps, landmark localisation requires selecting a suitable object detection model to locate the positions of the spine and femoral heads quickly and accurately. Parameter measurements require adding auxiliary lines according to the positions of keypoints and using geometric theory to solve the value of the spinopelvic parameter. The details are elaborated as follows.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eRadiography dataset preprocessing\u003c/h2\u003e \u003cp\u003eIt is necessary to preprocess the X-ray image to meet the input requirements of the landmark location model, as depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e (a). Preprocessing comprised two parts: cropping and resizing. The bottom half (1824\u0026times;1932) was derived from the full X-ray image and resized to 512\u0026times;512. Data labelling is another important task. All X-ray images were manually labelled for key landmarks to generate the ground truth for model training. A detailed illustration of the location of the four key landmarks on a radiograph and corresponding information are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e (b). The two keypoints\u0026mdash;S1L and S1C\u0026mdash;represent the leading edge and centre of the S1 endplate. The other two keypoints\u0026mdash;FH1 and FH2\u0026mdash;define the centre of the femoral head. The data labelling task was completed by two surgeons using LabelMe software version 5.1.1 (MIT, Boston, Massachusetts, MA, USA) and confirmed by an expert, which guaranteed the accuracy of keypoint labelling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eLandmark localisation model\u003c/h2\u003e \u003cp\u003eCompared with the existing research, the landmark location model based on CenterNet[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. In this paper a one-stage anchor-free object detection method is used that achieves a trade-off between accuracy and speed. The architecture of the Landmark Location model comprises two parts: the backbone and the output heads (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Using DLA-34[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] as the network backbone for the model, four output heads predicted the centre heatmap, centre offset, corner heatmap, and corner offset. The result of the two offset heads was used to compensate for the coordinate deviation caused by down-sampling rounding. The trained model could automatically infer the positions of the four keypoints in a radiograph that had not been marked with keypoints (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe loss function included heatmap prediction loss and offset regression loss. The former, an improved version of focal loss, measured the disparity between the predicted heatmap and the actual target location. The latter utilised L1 norm loss to measure the difference between the predicted and actual offsets. Meanwhile, data augmentation technology was implemented during model training to improve the model's accuracy and reduce data overfitting. Variations in brightness, contrast, and sharpness may be present in radiographs obtained from different machines, which can impact the precision and generalisability of the model. Therefore, data augmentation techniques were employed to effectively tackle this problem. We applied various modifications (such as rotated, flipped, and cropped) to the radiograph to simulate the variations in angle, direction, and exposure time during image acquisition. Specifically, we rotated, flipped, and cropped the images. In addition, we adjusted the contrast and brightness of the images to simulate differences in exposure time and added noise to simulate the variations in the acquisition process.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eParameter measurement\u003c/h2\u003e \u003cp\u003eSpinopelvic parameters can be calculated from the four predicted key landmarks by adding appropriate auxiliary lines. We fitted two lines and made four additional lines based on the four key landmarks, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e (c): one was the connecting line between the centres of the two femoral heads (whose midpoint was called the hip axis [HA]), and the other was along the upper edge of the sacral (S1) endplate; the four auxiliary lines included a horizontal reference, a vertical reference, a line connecting the HA with the centre of the S1 endplate, and a line orthogonal to the S1 endplate. Finally, we measured the three spinopelvic parameters PI, PT, and SS using geometric relationships among the six lines.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel training\u003c/h2\u003e \u003cp\u003eThe dataset comprised the training (800 images) and test (200 images) sets. The former was used to train the model using the Pytorch framework[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] on an NVIDIA GeForce GTX 3090Ti GPU for fitting the parameters of the model. The latter was used to evaluate the performance of the model by comparing its results with manually measured results. The model was initialised with publicly available pre-trained weights, the learning rate was set to 0.001, and the optimiser chose Adam. During training, the model was trained for 100 epochs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eOur method was evaluated using the test dataset (200 images) for landmark localisation error and spinopelvic parameter estimation. Localisation error refers to the difference between the predicted position of a keypoint and the actual position annotated by the expert. The test set was fed into the landmark localisation model to predict the position coordinates of four keypoints. Since the local errors were non-normally distributed, we used a boxen plot to visualise error distributions.\u003c/p\u003e \u003cp\u003eTo evaluate the intra- and inter-evaluator reliability of manual and model-based measurements of spinopelvic parameters, we enlisted two resident physicians (Evaluators 1 and 2) and one attending surgeon (Evaluator 3) to measure the spinopelvic parameters in the test dataset, with Evaluators 1 and 2 performing one measurement each, and Evaluator 3 conducting two measurements (Evaluator 3a and 3b), which were used as the gold standard for comparison. The angles of spinopelvic parameters in the test set were measured using the SurgiMap Spine software (Nemaris, Methuen, Massachusetts, MA, USA). All measurements were performed independently and blindly, meaning that after Evaluator 3 completed the first measurement, a second measurement was performed two weeks later. Moreover, the results of one evaluator were not eligible for a review by another evaluator. These manually measured values were used as the ground truth to compare with those obtained using our method (AI method) predictions and hence validate the algorithm.\u003c/p\u003e \u003cp\u003eFor spinopelvic parameter estimation, we employed four evaluation strategies: mean absolute difference (MAD) and its corresponding standard deviation (SD), the mean of errors between the measurement and the gold standard (ME), intraclass correlation coefficient (ICC), and Pearson correlation coefficient (R)[\u003cspan additionalcitationids=\"CR23\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. ICCs were used to evaluate the agreement between the model- and expert-measured values. ICCs between 0.75 and 1.00 were interpreted as having excellent reliability (good, 0.60\u0026ndash;0.74; fair, 0.40\u0026ndash;0.59; and poor, \u0026lt;\u0026thinsp;0.40)[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Pearson correlation coefficients were used to evaluate the correlations between the predicted spinopelvic parameters and their corresponding ground truth values. Bland\u0026ndash;Altman (B\u0026amp;A) plots were used to estimate the consistency intervals between the expert- and model-measured results[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. In this study, all statistical analyses were conducted using Python version 3.3.6.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003ePerformance of landmark localisation\u003c/h2\u003e \u003cp\u003eThe localisation errors of the keypoints are visualised using the boxen plots in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The localisation error of the centre of the S1 endplate (S1C) was the smallest (median error, 1.74 mm), and the median localisation error for the anterior edge of the S1 endplate (S1L) was approximately 2.03 mm, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a). When abnormal samples were removed from the test set, we observed that the landmark location error of the anterior edge on S1L was the smallest among the four landmarks (median error, 2.28 mm), as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eParameter measurement result\u003c/h2\u003e \u003cp\u003eThe intra- and inter-evaluator reliability analyses of manual measurements are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. ICCs for the inter-evaluator reliability were always the smallest for SS (0.945, Evaluator 1 vs Evaluator 3a) and the largest for PT (0.996, Evaluator 2 vs Evaluator 3a). R values were the largest for PT (0.995, Evaluator 2 vs Evaluator 3a); MAD values were also the largest for PT (2.143, Evaluator 2 vs Evaluator3a). The ME values were smaller than 2.76\u0026deg;, with the largest values for SS (Evaluator 1 vs Evaluator 3a). For the inter-evaluator reliability, ICCs were the smallest for SS and largest for PT. Additionally, all MAD values were less than 0.65.\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\u003eIntra- and inter-evaluator reliability of the manual measurements for spinopelvic parameters\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStatistical method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSS\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"18\" rowspan=\"19\"\u003e \u003cp\u003eInter-evaluator reliability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eEvaluator 1\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.591 (1.104)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.829 (0.582)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.571 (1.039)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.63\u0026thinsp;\u0026plusmn;\u0026thinsp;1.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.56\u0026thinsp;\u0026plusmn;\u0026thinsp;1.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.56\u0026thinsp;\u0026plusmn;\u0026thinsp;1.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.943 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-87}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.985 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-138}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.922 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-75}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.965 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-92}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.995 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-168}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.955 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-83}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eEvaluator 1\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.273 (1.075)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.500 (0.429)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.313 (1.088)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.19\u0026thinsp;\u0026plusmn;\u0026thinsp;1.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.46\u0026thinsp;\u0026plusmn;\u0026thinsp;0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.12\u0026thinsp;\u0026plusmn;\u0026thinsp;1.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.959 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-85}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.996 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-172}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.949 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-78}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.964 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-91}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.994 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-160}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.954(\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eEvaluator 2\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.582 (1.290)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.425 (0.295)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.646(1.210)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.62\u0026thinsp;\u0026plusmn;\u0026thinsp;2.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.76\u0026thinsp;\u0026plusmn;\u0026thinsp;0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.76\u0026thinsp;\u0026plusmn;\u0026thinsp;2.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.943 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-86}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.995 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-179}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.934 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-80}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.960 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-86}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.996 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-180}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.952 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-79}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eEvaluator 2\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.423 (1.254)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.379 (0.301)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.401 (1.132)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.39\u0026thinsp;\u0026plusmn;\u0026thinsp;1.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.47\u0026thinsp;\u0026plusmn;\u0026thinsp;0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.46\u0026thinsp;\u0026plusmn;\u0026thinsp;1.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.941 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-85}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.994 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-171}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.930 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-78}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.959 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-85}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.996 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-172}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.949 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-78}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eEvaluator 1\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.871 (1.346)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.859 (0.591)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.853 (1.372)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.933 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-80}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.984 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-134}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.920 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-73}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.951 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-79}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.989 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-135}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.941 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-72}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eIntra-evaluator reliability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eEvaluator 3a\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.450 (0.409)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.374 (0.305)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.611 (0.607)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.996 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-190}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.997 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-200}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.991 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-160}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.997 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-192}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.998 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-201}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.994 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-162}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eMAD, mean absolute difference; ME, mean error; SD, standard deviation; ICC, intraclass correlation coefficient; PI, pelvic incidence; PT, pelvic tilt; SS, sacral slope\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFor the inter-evaluator reliability, analysis between the AI model and manual measurements was conducted, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. ICC values for spinopelvic parameters ranged between 0.919 and 0.997, with the smallest values for SS (AI method vs Evaluator 2) and the largest for PT (AI method vs Evaluator 3a). The MAD values were the smallest for PT (0.367, Evaluator 3a vs AI) and the largest for SS (2.097, Evaluator 2 vs AI method). The R values for spinopelvic parameters were the smallest for SS (0.89, evaluator vs AI).\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\u003eInter-evaluator reliability of the manual measurements and AI method for spinopelvic parameters\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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 \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStatistical method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSS\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eAI method\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.318 (1.180)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.978(0.613)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.354 (1.167)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.22\u0026thinsp;\u0026plusmn;\u0026thinsp;1.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.18\u0026thinsp;\u0026plusmn;\u0026thinsp;1.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.98\u0026thinsp;\u0026plusmn;\u0026thinsp;1.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.948 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-90}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.982 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-131}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.924 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-76}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.964 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-91}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.987 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-132}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.924 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-77}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eAI method\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.931 (1.263)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.437 (0.275)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2.097 (1.324)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.40\u0026thinsp;\u0026plusmn;\u0026thinsp;2.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3.97\u0026thinsp;\u0026plusmn;\u0026thinsp;2.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.931 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-78}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.992 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-162}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.899 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-72}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.949 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-77}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.994 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-163}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.919 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-67}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eAI method\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.941 (0.772)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.367 (0.299)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.035 (0.753)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.47\u0026thinsp;\u0026plusmn;\u0026thinsp;1.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.62\u0026thinsp;\u0026plusmn;\u0026thinsp;0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.74\u0026thinsp;\u0026plusmn;\u0026thinsp;1.28\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.976 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-120}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.995 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-187}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.965 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-106}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.983 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-120}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.997 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-188}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.975 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-105}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eAI method\u003c/p\u003e \u003cp\u003evs\u003c/p\u003e \u003cp\u003eEvaluator 3b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAD (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.975 (0.793)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.376 (0.299)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.059 (0.814)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.50\u0026thinsp;\u0026plusmn;\u0026thinsp;1.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1.86\u0026thinsp;\u0026plusmn;\u0026thinsp;1.33\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR (\u003cem\u003ep\u003c/em\u003e-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.975 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-119}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.995 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-186}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.961 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-101}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eICC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.983 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-119}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.997 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-187}\\)\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.972 (\u0026lt;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-100}\\)\u003c/span\u003e\u003c/span\u003e)\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\u003ePI, pelvic incidence; PT, pelvic tilt; SS, sacral slope; AI, artificial intelligence\u003c/p\u003e \u003cp\u003eCompared to manual measurements, the AI method showed the smallest errors for PI, TT, and SS when compared to Evaluator 3a (gold standard), with values of 0.16, 0.27, and 1.13, respectively, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Exemplarily, the correlation and agreement between AI and Evaluator 3a for spinopelvic parameters are further depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e7\u003c/span\u003e presents the detection success rate of the parameters, which refers to the proportion of test cases with the measured MAD lower than the threshold. Specifically, the AI-measured results were close to the mean of the expert manual measurement results while exhibiting a smaller variance. The evaluator in this study reported an average time of 2 to 3 minutes for measuring spinopelvic parameters per image, whereas the proposed algorithm accomplished the same analysis in approximately 5 seconds.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAngle measurement by the manual measurements and AI method for spinopelvic parameters\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSS\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAI Method\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e45.76\u0026deg;\u0026plusmn;8.00\u0026deg; (29.60\u0026deg;, 60.96\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e10.03\u0026deg;\u0026plusmn;6.41\u0026deg; (-2.52\u0026deg;, 22.58\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e35.26\u0026deg;\u0026plusmn;6.61\u0026deg; (22.58\u0026deg;, 48.44\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluator 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e44.05\u0026deg;\u0026plusmn;7.81\u0026deg; (28.26\u0026deg;, 61.57\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e7.88\u0026deg;\u0026plusmn;6.45\u0026deg; (-2.53\u0026deg;, 22.90\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e34.62\u0026deg;\u0026plusmn;6.71\u0026deg; (21.48\u0026deg;, 47.76\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluator 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e47.75\u0026deg;\u0026plusmn;8.58\u0026deg; (30.92\u0026deg;, 64.57\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e10.44\u0026deg;\u0026plusmn;6.34\u0026deg; (-1.98\u0026deg;, 22.85\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e38.71\u0026deg;\u0026plusmn;7.57\u0026deg; (23.86\u0026deg;, 53.54\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluator 3a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e45.92\u0026deg;\u0026plusmn;8.33\u0026deg; (29.59\u0026deg;, 62.24\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e9.52\u0026deg;\u0026plusmn;6.37\u0026deg; (-2.97\u0026deg;, 22.00\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e36.39\u0026deg;\u0026plusmn;7.02\u0026deg; (22.62\u0026deg;, 50.15\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluator 3b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e46.03\u0026deg;\u0026plusmn;8.27\u0026deg; (27.63\u0026deg;, 60.20\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e9.52\u0026deg;\u0026plusmn;6.38\u0026deg; (-2.97\u0026deg;, 22.02\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e36.52\u0026deg;\u0026plusmn;7.02\u0026deg; (22.76\u0026deg;,50.26\u0026deg;)\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\u003ePI, pelvic incidence; PT, pelvic tilt; SS, sacral slope\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eSagittal balance plays a vital role in the diagnosis, treatment, and surgical planning of spinal disorders. We presented an automated method to measure pelvic parameters and compared it with expert manual measurements. Excellent agreement was demonstrated between human measurement and the automated method for the spinopelvic parameters (PI, PT, and SS).\u003c/p\u003e \u003cp\u003eIt can be observed that the localisation errors of F1 and F2 are larger than those of the anterior edge and the centre of S1 in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003e (b). As an X-ray is taken from the lateral spine, there was an inevitable overlap of the femoral head. The distant femoral head is easily occluded, resulting in the false marking of keypoints. The smaller the landmark error, the closer the position of the keypoint is to the real position, and the more accurate the calculated parameter value. Overall, the localisation error of the four key landmarks was \u0026lt;\u0026thinsp;8.0 mm for most X-rays in the test set, an acceptable error range for expert surgeon measurements.\u003c/p\u003e \u003cp\u003eThe measurement of parameters in spinal imaging is time-consuming and particularly prone to inconsistency. Therefore, several studies have evaluated intraobserver and interobserver variability in sagittal spinopelvic balance parameters measured manually from radiographs with computer-aided tools[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Malioot et al. reported an SD of 3.5\u0026deg; when measuring SS, PT, and PI using the Keops software (SMAIO, Lyon, France)[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Lafage et al. evaluated the efficacy of the SurgiMap Spine software (Nemaris, Methuen, Massachusetts, MA, USA) and reported intraobserver standard deviations of 4.6\u0026deg;, 2.5\u0026deg;, 0.8\u0026deg;, and 5.4\u0026deg; for SS, PT, and an angle similar to ST and PI, respectively[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Vila-Casademunt et al. also evaluated the SurgiMap Spine software by measuring intraobserver MAD and interobserver MAD for spinopelvic balance parameters. They observed that the presence of lumbosacral instrumentation resulted in a significant increase in interobserver variability[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Although parameter measurement errors can be reduced using computer-aided tools, they are still classified as manual measurements owing to the use of the mouse to mark points and draw lines and other geometric shapes, including circles, on the X-ray. The measurement of pelvic parameters continued to be affected by the X-ray quality and observer experience, and the validity and reliability of parameter measurement could not be guaranteed.\u003c/p\u003e \u003cp\u003eThe emergence of AI has made automated measurement tools more intelligent and automated, which not only helps avoid repetitive manual measurement but also saves the time cost in clinical practice. Galbusera et al. proposed a DL model based on the location of keypoints for spinopelvic parameter prediction[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], which calculated the difference between AI and the manual measurements of PT, PI, and SS, yielding MADs (SD) of is 2.7\u0026deg; (0.7\u0026deg;), 9.5\u0026deg; (8.5\u0026deg;), and 6.9\u0026deg; (6.2\u0026deg;), respectively. Korez et al. proposed a two-stage model based on the U-Net network structure to implement a DL measurement tool for the fully automated measurement of spinopelvic parameters[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The MADs (SD) between manual measurements and automated tools for PT, PI, and SS were 2.7\u0026deg; (2.5\u0026deg;), 1.2\u0026deg; (1.2\u0026deg;), and 5.0\u0026deg; (3.4\u0026deg;), respectively. Furthermore, decentralised convolutional DL-based approaches have been proven to significantly improve parameter measurement accuracy. Chae et al. presented an automated method for measuring spinopelvic parameters with reported MADs of 1.45\u0026deg;, 2.51\u0026deg;, and 2.64\u0026deg; for PT, PI, and SS, respectively[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. The method proposed by Nguyen et al. had MADs (SD) of 1.16\u0026deg; (0.82\u0026deg;), 2.21\u0026deg; (2.05\u0026deg;), and 3.17\u0026deg; (3.09\u0026deg;) for PT, PI, and SS, respectively, representing the lowest deviations compared with other methods reported in the previous literature[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOur method enables the fully automatic measurement of spinopelvic parameters using a simple one-stage DL model. Compared to other methods, this approach achieves higher efficiency by directly performing one-stage detection through keypoint detection. It shortens the measurement time through a single training and eliminates the need for multi-stage detection. More importantly, our method achieves better inter-class agreement with manual measurements.\u003c/p\u003e \u003cp\u003eTo validate the measurement quality of human evaluators in evaluating the AI algorithm, intra- and inter-evaluator measurements were performed. PT showed the highest reliability among the spinopelvic parameters between human evaluators (Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), while SS had relatively lower ICCs values. The results indicate high reliability within the three evaluators (ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.93), which is well in line with previously published assessments[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] and demonstrates the validity of manual measurements as the ground truth to evaluate AI algorithms.\u003c/p\u003e \u003cp\u003eIt proved the reliability of measuring parameters on radiographic images for the one-stage DL model. The agreement between the AI method and one attending surgeon's measurements was higher than that between one attending surgeon and two resident physicians' measurements. For PT, PI, and SS, the ICCs and the R value between AI and Evaluator 3a, as well as Evaluator 3b, were higher when compared to the corresponding values derived from Evaluator 1 and Evaluator 2. Moreover, the MAD values between AI and Evaluator 3a, as well as Evaluator 3b, were smaller in comparison to the values computed between Evaluator 1 and Evaluator 2, as evident from Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn the meantime, we analysed the consistency between AI and Evaluator 3a (gold standard). The ICCs ranged between 0.975 and 0.997. \u003cem\u003eR\u003c/em\u003e was between 0.965 and 0.995, and MAD (SD) was 0.941\u0026deg; (0.772\u0026deg;) for PI, 0.367\u0026deg; (0.299\u0026deg;) for PT, and 1.035\u0026deg; (0.753\u0026deg;) for SS in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Exemplarily, the MAD (SD) for AI vs Evaluator 3a was lower than the values for instance reported by Nguyen et al.[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] for the spinopelvic parameters (PI 0.941\u0026deg;\u0026lt;2.21\u0026deg;; PT 0.367\u0026deg;\u0026lt;1.16\u0026deg;; SS 1.035\u0026deg;\u0026lt;3.17\u0026deg;). Moreover, the correlation scatter diagrams in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e6\u003c/span\u003e (a), (c), and (e) and the B\u0026amp;A diagrams in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e6\u003c/span\u003e (b), (d), and (f) further showed that the results measured using our method were in good agreement with the results marked by the attending surgeon. Compared to the other two parameters, the consistency of AI and expert measurement in measuring SS was slightly lower, but still within an acceptable margin of error for experts. As demonstrated in the B\u0026amp;A diagrams, the differences in measurement results followed a normal distribution; 95% of the differences should fall within the MD\u0026thinsp;\u0026plusmn;\u0026thinsp;1.96 SD interval, known as the 95% limits of agreement (95% LoA), with only a few points out of bounds. This indicates that the AI-based measurement method is highly consistent with the results obtained by human evaluator measurements.\u003c/p\u003e \u003cp\u003eOur method had a higher measurement accuracy and shorter time overhead. Compared to the measurements obtained by two resident physicians, the measurement accuracy of the AI method was closer to that obtained by one attending surgeon on two separate occasions in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The measurement time for all spinopelvic parameters in each case using this method has been reduced by approximately 30 times compared to manual measurements, resulting in a significant improvement in the detection speed.\u003c/p\u003e \u003cp\u003eFurthermore, we analysed the detection success rate, which means that when the error angle was smaller than the threshold, the detection value was considered acceptable. In Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e7\u003c/span\u003e, when the degree threshold was set to 3\u0026deg;, the detection success rate exceeded 80%. Experts can recognise and accept this threshold angle. Of note, this threshold angle fell within an acceptable margin of error for experts and did not have any impact on the final diagnostic results.\u003c/p\u003e \u003cp\u003eOur approach demonstrated higher reliability compared to other methods, as evidenced by the statistical analysis conducted using manual measurements from multiple experts. When compared to expert manual measurements, our method exhibited superior detection accuracy. The evaluation of measurement time validated the higher measurement speed of our method. Moreover, we collected a larger number of X-ray data samples as the training and testing sets. Increasing the number of X-ray data samples for the training dataset could improve the AI model in extracting rich feature information to accurately identify keypoints and calculate the pelvic parameters. Additionally, a larger number of test data samples could help ensure the fairness of the model evaluation process and verify the generalisation capabilities of the AI model.\u003c/p\u003e \u003cp\u003eHowever, this study has some limitations. First, experts still exhibited certain biases in making judgements based on small amounts of complex data. Second, the model could not effectively correct the errors caused by states such as coronal deformities and anatomical differences, resulting in measurement errors. In the future, we intend to leverage knowledge transfer techniques from the field of AI to acquire feature knowledge from complex samples, thereby reducing measurement errors.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eOur study proposed a new method for the fully automatic measurement of PI, PT, and SS spinopelvic parameters via the DL model-driven identification of key landmarks. The statistical analyses were built based on the deviation of the proposed model predictions versus measurements by three evaluators who conducted manual measurement using a test dataset of 1,000 radiographs. Our method has been proven to possess high measurement accuracy, excellent consistency, and faster measurement speed for spinopelvic parameter measurement. It effectively alleviated laborious routine tasks, saving valuable time and reducing the likelihood of error-prone manual measurements. Furthermore, its potential applicability extends to other orthopaedic sub-domains.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Artificial intelligence\u003c/p\u003e\n\u003cp\u003eDL\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Deep learning\u003c/p\u003e\n\u003cp\u003eHA\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Hip axis\u003c/p\u003e\n\u003cp\u003eICCV\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;International Conference on Computer Vision\u003c/p\u003e\n\u003cp\u003eMD\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Mean difference\u003c/p\u003e\n\u003cp\u003eSD\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Standard deviation\u003c/p\u003e\n\u003cp\u003eICC\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Intraclass correlation coefficient\u003c/p\u003e\n\u003cp\u003eMAD\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Mean absolute difference\u003c/p\u003e\n\u003cp\u003ePI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Pelvic incidence\u003c/p\u003e\n\u003cp\u003ePT \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Pelvic tilt\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the ethics committee of China Beijing Chaoyang hospital, Capital Medical Univercity and the requirement for informed consent was waived due to its retrospective nature. \u0026nbsp; The approval document number is 2021-8-3-1. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors (Xianglong Meng, Jianhua Liu, Zihe Feng , Yu Sun, Zhiheng Zhao, Zhiqiang Bai, and Yong Hai) declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the R\u0026amp;D Program of Beijing Municipal Education Commission (NO. KZ202210025034), the Science and Technology and Information Technology Bureau of Chaoyang District, Beijing (CN) (NO. CYSF2019), and the National Natural Science Foundation of China ((NO. 6227070318).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e* Xianglong Meng,\u0026nbsp;Jianhua Liu, and Zihe Feng 1 contributed equally to this study.\u003c/p\u003e\n\u003cp\u003eConceptualization: Xianglong Meng, Yu Sun; Methodology: Jianhua Liu, Zihe Feng ; Formal analysis and investigation: Jianhua Liu, Zihe Feng, Zhiheng Zhao; Writing-original draft preparation: Jianhua Liu, Zihe Feng ; Writing-review and editing: Xianglong Meng, Yu Sun, Jianhua Liu, Zihe Feng ; Supervision: Xianglong Meng, Yu Sun, Zhiqiang Bai, Yong Hai. All authors approved the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAzimi P, Yazdanian T, Benzel EC, Hai Y, Montazeri A: \u003cstrong\u003eSagittal balance of the cervical spine: A systematic review and meta-analysis\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2021, \u003cstrong\u003e30\u003c/strong\u003e:1411-1439.\u003c/li\u003e\n\u003cli\u003eSchl\u0026ouml;sser TP, Castelein RM, Grobost P, Shah SA, Abelin-Genevois K: \u003cstrong\u003eSpecific sagittal alignment patterns are already present in mild adolescent idiopathic scoliosis\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2021, \u003cstrong\u003e30\u003c/strong\u003e:1881-1887.\u003c/li\u003e\n\u003cli\u003eVrtovec T, Janssen MM, Likar B, Castelein RM, Viergever MA, Pernu\u0026scaron; F: \u003cstrong\u003eA review of methods for evaluating the quantitative parameters of sagittal pelvic alignment\u003c/strong\u003e. \u003cem\u003eSpine J \u003c/em\u003e2012, \u003cstrong\u003e12\u003c/strong\u003e(5):433-446.\u003c/li\u003e\n\u003cli\u003eKorez R, Putzier M, Vrtovec T: \u003cstrong\u003eAutomated measurement of pelvic incidence from X-ray images\u003c/strong\u003e. In\u003cem\u003e.\u003c/em\u003e Cham: Springer International Publishing; 2019: 146-152.\u003c/li\u003e\n\u003cli\u003eKorez R, Putzier M, Vrtovec T: \u003cstrong\u003eComputer-assisted measurement of sagittal pelvic alignment parameters from radiographic images\u003c/strong\u003e. \u003cem\u003eZdravniski Vestnik \u003c/em\u003e2018, \u003cstrong\u003e87\u003c/strong\u003e(11-12):519-529.\u003c/li\u003e\n\u003cli\u003eGarbossa D, Pejrona M, Damilano M, Sansone V, Ducati A, Berjano P: \u003cstrong\u003ePelvic parameters and global spine balance for spine degenerative disease: the importance of containing for the well being of content\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2014, \u003cstrong\u003e23\u003c/strong\u003e(6):616-627.\u003c/li\u003e\n\u003cli\u003eKrupinski EA: \u003cstrong\u003eCurrent perspectives in medical image perception\u003c/strong\u003e. \u003cem\u003eAtten Percept Psychophys \u003c/em\u003e2010, \u003cstrong\u003e72\u003c/strong\u003e(5):1205-1217.\u003c/li\u003e\n\u003cli\u003eMaillot C, Ferrero E, Fort D, Heyberger C, Le Huec JC: \u003cstrong\u003eReproducibility and repeatability of a new computerized software for sagittal spinopelvic and scoliosis curvature radiologic measurements: Keops\u0026reg;\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2015, \u003cstrong\u003e24\u003c/strong\u003e(7):1574-1581.\u003c/li\u003e\n\u003cli\u003eLafage R, Ferrero E, Henry JK, Challier V, Diebo B, Liabaud B, Lafage V, Schwab F: \u003cstrong\u003eValidation of a new computer-assisted tool to measure spino-pelvic parameters\u003c/strong\u003e.\u003cem\u003e Spine J \u003c/em\u003e2015, \u003cstrong\u003e15\u003c/strong\u003e(12):2493-2502.\u003c/li\u003e\n\u003cli\u003eCharles YP, Lamas V, Ntilikina Y: \u003cstrong\u003eArtificial intelligence and treatment algorithms in spine surgery\u003c/strong\u003e. \u003cem\u003eOrthop Traumatol Surg Res \u003c/em\u003e2023, \u003cstrong\u003e109\u003c/strong\u003e(1, Supplement):103456.\u003c/li\u003e\n\u003cli\u003eAzimi P, Yazdanian T, Benzel EC, Aghaei HN, Azhari S, Sadeghi S, Montazeri A: \u003cstrong\u003eA review on the use of artificial intelligence in spinal diseases\u003c/strong\u003e. \u003cem\u003eAsian Spine J \u003c/em\u003e2020, \u003cstrong\u003e14\u003c/strong\u003e(4):543.\u003c/li\u003e\n\u003cli\u003eAubert B, Vazquez C, Cresson T, Parent S, Guise JAd: \u003cstrong\u003eToward automated 3D spine reconstruction from biplanar radiographs using CNN for statistical spine model fitting\u003c/strong\u003e. \u003cem\u003eIEEE Trans Med Imaging \u003c/em\u003e2019, \u003cstrong\u003e38\u003c/strong\u003e(12):2796-2806.\u003c/li\u003e\n\u003cli\u003eOrosz L, Haines CM, Thomson A, Schuler TC, Good CR, Grover P, Dreischarf M, Roy R, Jazini E: \u003cstrong\u003e74. Novel artificial intelligence algorithm can accurately and independently measure spinopelvic parameters\u003c/strong\u003e. \u003cem\u003eSpine J \u003c/em\u003e2021, \u003cstrong\u003e21\u003c/strong\u003e:S36-S37.\u003c/li\u003e\n\u003cli\u003eGalbusera F, Niemeyer F, Wilke H-J, Bassani T, Casaroli G, Anania C, Costa F, Brayda-Bruno M, Sconfienza LM: \u003cstrong\u003eFully automated radiological analysis of spinal disorders and deformities: a deep learning approach\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2019, \u003cstrong\u003e28\u003c/strong\u003e(5):951-960.\u003c/li\u003e\n\u003cli\u003eWu H, Bailey C, Rasoulinejad P, Li S: \u003cstrong\u003eAutomated comprehensive adolescent idiopathic scoliosis assessment using MVC-Net\u003c/strong\u003e. \u003cem\u003eM Image Anal \u003c/em\u003e2018, \u003cstrong\u003e48\u003c/strong\u003e:1-11.\u003c/li\u003e\n\u003cli\u003eWeng C-H, Wang C-L, Huang Y-J, Yeh Y-C, Fu C-J, Yeh C-Y, Tsai T-T: \u003cstrong\u003eArtificial intelligence for automatic measurement of sagittal vertical axis using ResUNet framework\u003c/strong\u003e. \u003cem\u003eJ Clin Med \u003c/em\u003e2019, \u003cstrong\u003e8\u003c/strong\u003e(11):1826.\u003c/li\u003e\n\u003cli\u003eKorez R, Putzier M, Vrtovec T: \u003cstrong\u003eA deep learning tool for fully automated measurements of sagittal spinopelvic balance from X-ray images: performance evaluation\u003c/strong\u003e. \u003cem\u003eEur Spine J \u003c/em\u003e2020, \u003cstrong\u003e29\u003c/strong\u003e(9):2295-2305.\u003c/li\u003e\n\u003cli\u003eCho BH, Kaji D, Cheung ZB, Ye IB, Tang R, Ahn A, Carrillo O, Schwartz JT, Valliani AA, Oermann EK\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eAutomated measurement of lumbar lordosis on radiographs using machine learning and computer vision\u003c/strong\u003e. \u003cem\u003eGlobal Spine J \u003c/em\u003e2020, \u003cstrong\u003e10\u003c/strong\u003e(5):611-618.\u003c/li\u003e\n\u003cli\u003eDuan K, Bai S, Xie L, Qi H, Huang Q, Tian Q: \u003cstrong\u003eCenterNet: Keypoint Triplets for Object Detection\u003c/strong\u003e. In: \u003cem\u003e2019 IEEE/CVF International Conference on Computer Vision.\u003c/em\u003e IEEE; 2019: 6568-6577.\u003c/li\u003e\n\u003cli\u003eZhou X, Koltun V, Kr\u0026auml;henb\u0026uuml;hl P: \u003cstrong\u003eTracking Objects as Points\u003c/strong\u003e. In: \u003cem\u003eEuropean Conference on Computer Vision.\u003c/em\u003e Cham: Springer; 2020: 474-490.\u003c/li\u003e\n\u003cli\u003eChoi J, Yeom HY, Kim Y: \u003cstrong\u003eImproving Oversubscribed GPU Memory Performance in the PyTorch Framework\u003c/strong\u003e. \u003cem\u003eCluster Computing \u003c/em\u003e2022.\u003c/li\u003e\n\u003cli\u003eCicchetti DV: \u003cstrong\u003eGuidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology\u003c/strong\u003e. \u003cem\u003ePsychological Assessment \u003c/em\u003e1994, \u003cstrong\u003e6\u003c/strong\u003e:284-290.\u003c/li\u003e\n\u003cli\u003eWEIR JP: \u003cstrong\u003eQuantifying test-retest reliability using the intraclass correlation coefficient and the SEM\u003c/strong\u003e. \u003cem\u003eThe Journal of Strength \u0026amp; Conditioning Research \u003c/em\u003e2005, \u003cstrong\u003e19\u003c/strong\u003e(1):231-240.\u003c/li\u003e\n\u003cli\u003eCohen I, Huang Y, Chen J, Benesty J, Benesty J, Chen J, Huang Y, Cohen I: \u003cstrong\u003ePearson correlation coefficient\u003c/strong\u003e. \u003cem\u003eNoise reduction in speech processing \u003c/em\u003e2009:1-4.\u003c/li\u003e\n\u003cli\u003eHarrison DE, Cailliet R, Harrison DD, Janik TJ, Holland B: \u003cstrong\u003eReliability of Centroid, Cobb, and Harrison Posterior Tangent Methods: Which to Choose for Analysis of Thoracic Kyphosis\u003c/strong\u003e. \u003cem\u003eSpine \u003c/em\u003e2001, \u003cstrong\u003e26\u003c/strong\u003e(11):e227-e234.\u003c/li\u003e\n\u003cli\u003eAltman DG, Bland JM: \u003cstrong\u003eMeasurement in medicine: the analysis of method comparison studies\u003c/strong\u003e. \u003cem\u003eJ R Stat Soc Ser A Stat Soc: Series D \u003c/em\u003e1983, \u003cstrong\u003e32\u003c/strong\u003e(3):307-317.\u003c/li\u003e\n\u003cli\u003eVila-Casademunt A, Pellis\u0026eacute; F, Acaroglu E, P\u0026eacute;rez-Grueso FJS, Mart\u0026iacute;n-Buitrago MP, Sanli T, Yakici S, de Frutos AG, Matamalas A, S\u0026aacute;nchez-M\u0026aacute;rquez JM\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eThe reliability of sagittal pelvic parameters: the effect of lumbosacral instrumentation and measurement experience\u003c/strong\u003e. \u003cem\u003eSpine \u003c/em\u003e2015, \u003cstrong\u003e40\u003c/strong\u003e(4):E253-E258.\u003c/li\u003e\n\u003cli\u003eChae D-s, Nguyen TP, Park S-J, Kang K-Y, Won C, Yoon J: \u003cstrong\u003eDecentralized convolutional neural network for evaluating spinal deformity with spinopelvic parameters\u003c/strong\u003e. \u003cem\u003eComput Methods Programs Biomed \u003c/em\u003e2020, \u003cstrong\u003e197\u003c/strong\u003e:105699.\u003c/li\u003e\n\u003cli\u003eNguyen TP, Jung JW, Yoo YJ, Choi SH, Yoon J: \u003cstrong\u003eIntelligent evaluation of global spinal alignment by a decentralized convolutional neural network\u003c/strong\u003e. \u003cem\u003eJ Digit Imaging \u003c/em\u003e2022, \u003cstrong\u003e35\u003c/strong\u003e(2):213-225.\u003c/li\u003e\n\u003cli\u003eSomoske\u0026ouml;y S, Tunyogi-Csap\u0026oacute; M, Bogy\u0026oacute; C, Ill\u0026eacute;s T: \u003cstrong\u003eAccuracy and reliability of coronal and sagittal spinal curvature data based on patient-specific three-dimensional models created by the EOS 2D/3D imaging system\u003c/strong\u003e. \u003cem\u003eSpine J \u003c/em\u003e2012, \u003cstrong\u003e12\u003c/strong\u003e(11):1052-1059.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"spinopelvic parameters, deep learning, landmark localisation, parameter measurement","lastPublishedDoi":"10.21203/rs.3.rs-3734310/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3734310/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eThe current method of measuring parameters in spinal imaging manually is time-consuming and prone to inconsistencies. This study proposed and validated a novel method to automate the measurement of pelvic parameters using a one-stage deep learning (DL) model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e Spinopelvic parameters, including pelvic incidence (PI), sacral slope (SS), and pelvic tilt (PT), were measured from full body radiographs of patients by three evaluators and by using our proposed method. Our proposed one-stage DL model was based on keypoint localisation. Landmark localisation error was used to evaluate the performance of landmark localisation. To evaluate the agreement between our method and the human evaluators, the analysis of average error, standard deviation, and intra- and inter-evaluator reliability was conducted using the intraclass correlation coefficient (ICC) and Pearson's correlation coefficient (\u003cem\u003eR\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003eThe method achieved excellent measurement performance for spinopelvic parameters. The distribution of the landmark localisation errors was within a reasonable range (median error, 2.28–4.01 mm). ICC values for the assessment of the intra- (range: 0.941–0.996) and inter-evaluator (0.994–0.998) reliability of human evaluators were excellent. The method was able to determine spinopelvic parameters with excellent ICC values (0.919-0.997) and \u003cem\u003eR\u003c/em\u003e value (\u003cem\u003eR \u003c/em\u003e\u0026gt;0.899, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001, all). Meanwhile, the detection speed of the algorithm was approximately 30 times faster than that of manual measurements of spinopelvic parameters.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003eThis one-step automated measurement method is less time-consuming and has excellent reliability and agreement with human evaluators.\u003c/p\u003e","manuscriptTitle":"Accurate automatic measurement of spinopelvic parameters with a one-stage deep learning technique","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-02 20:00:06","doi":"10.21203/rs.3.rs-3734310/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"041fe4a7-b293-434a-8184-19972ab64723","owner":[],"postedDate":"January 2nd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-11-23T08:38:48+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-02 20:00:06","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3734310","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3734310","identity":"rs-3734310","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}