A novel approach for screening standard anteroposterior pelvic radiographs in children

Anteroposterior pelvic radiography is the first‐line imaging modality for diagnosing developmental dysplasia of the hip (DDH). Nonstandard radiographs with pelvic malposition make the correct diagnosis of DDH challenging. However, as the only method available for screening standard pelvic radiographs, traditional manual assessment is relatively laborious and potentially erroneous. We retrospectively collected 3,247 pelvic radiographs. There were 2,887 radiographs randomly selected to train and optimize the AI model. Then 362 radiographs were used to test the model’s diagnostic performance. Its diagnostic accuracy was assessed using receiver operating characteristic (ROC) curves and measurement consistency using Bland–Altman plots. In 362 radiographs, the AI model’s area under ROC curves, accuracy, sensitivity, and specificity for quality assessment was 0.993, 99.4% (360/362), 98.6% (138/140), and 100.0% (222/222), respectively. Compared with clinicians, the 95% limits of agreement (Bland–Altman analysis) for pelvic tilt index (PTI) and pelvic rotation index (PRI), as determined by the model, were -0.052–0.072 and -0.088–0.055, respectively. Conclusions: The artificial intelligence-assisted method was more efficient and highly consistent with clinical experts. This method can be used for real-time validation of the quality of pelvic radiographs in current picture archiving and communications systems (PACS). What is Known: • Nonstandard pediatric radiographs with pelvic malposition make the correct diagnosis of developmental dysplasia of the hip (DDH) challenging. • Traditional manual assessment remains the only method available for screening standard pediatric pelvic radiographs, which is relatively laborious and potentially erroneous. What is New: • This study proposed an artificial intelligence-assisted model to assess the quality of pediatric pelvic radiographs accurately and efficiently. • We recommend the integration of the model into current picture archiving and communications systems (PACS) for real-time screening of standard pediatric pelvic radiographs. What is Known: • Nonstandard pediatric radiographs with pelvic malposition make the correct diagnosis of developmental dysplasia of the hip (DDH) challenging. • Traditional manual assessment remains the only method available for screening standard pediatric pelvic radiographs, which is relatively laborious and potentially erroneous. What is New: • This study proposed an artificial intelligence-assisted model to assess the quality of pediatric pelvic radiographs accurately and efficiently. • We recommend the integration of the model into current picture archiving and communications systems (PACS) for real-time screening of standard pediatric pelvic radiographs.


Introduction
Pediatric developmental dysplasia of the hip (DDH) is challenging to be examined clinically for its atypical symptoms at an early stage, relying on imaging modalities [1].Anteroposterior pelvic radiography is the first-line screening examination for diagnosing DDH in children over 4 to 6 months [2][3][4].Previous studies have confirmed that nonstandard radiographs with pelvic malposition significantly affect the measurements of hip parameters, thus leading to the misdiagnosis of DDH [5][6][7][8].However, standard pelvic radiographs are not always available in clinical practice owing to the high requirement for proper pelvic positioning and poor cooperation of children [9][10][11].Therefore, it is crucial to perform quality control of pelvic radiographs (i.e., determine whether they are standard) before diagnosing DDH [8].
Currently, the traditional manual evaluation remains the only method for screening standard pelvic radiographs, i.e., assessing the quality of pelvic radiographs [9,12,13].However, this method may have errors in estimating obturator diameters and be relatively laborious.The main steps were as follows: annotating anatomical key landmarks, drawing reference lines, estimating the obturator transverse and vertical diameters, calculating the pelvic rotation index (PRI) [12] and pelvic tilt index (PTI) [13], and making the final judgment according to the reference values.Some inexperienced clinicians even acquiesced in the assumption that all radiographs were standard, which may lead to misdiagnosis of DDH.To our knowledge, there was no quick and accurate method to assess the quality of pelvic radiographs.
In this study, we proposed and validated a convolutional neural network model to assess the quality of pediatric pelvic radiographs.The computer vision-based model can detect anatomical key landmarks and segment the obturator region on pelvic radiographs.Assuming great accuracy and efficiency in screening standard radiographs, which can be used for realtime validation of the quality of pelvic radiographs in current picture archiving and communications systems (PACS).

Patients
This retrospective study was approved by the local institutional ethics committee.The requirement for informed consent was waived due to the retrospective use of

Obturator morphology-based quality assessment for pelvic radiographs
The standard pelvic radiograph refers to a basically symmetrical size of bilateral obturator foramina, while the anterior pelvic plane is parallel to the coronal plane [9].Tönnis et al. and Ball et al. proposed the pelvic rotation index (PRI) and pelvic tilt index (PTI) to quantify pelvic positioning and screen standard pelvic radiographs [12,13].The PRI is obtained by dividing the right obturator transverse diameter (Rt) by the left one (Lt) (normal range, 0.56-1.80).The PTI is the ratio between the obturator vertical diameter (Dv) and the vertical distance of the symphysis pubis from Hilgenreiner's line (h), with reference values of 0.75 to 1.20.The ultimate measurement scheme for PRI and PTI is shown in Fig. 2.
The radiographic morphology of pediatric obturator foramina was classified according to the lower end of the ischial rami (defined as I-point): type 1: the I-point was lateral to the center obturator vertical line (COVL); type 2: the I-point was at COVL; type 3: the I-point was medial to COVL; and type 4: the lower ends of the ischial and pubic rami were fused (Fig. 3A-D).However, the morphological variation of obturator foramina in young children complicates the assessment method of standard pelvic radiographs.For type 1, the transverse diameter of the obturator foramen was parallel to the Hilgenreiner's line, passing the lower end of the pubic rami; its vertical diameter was perpendicular to the Hilgenreiner's line, passing through the lower end of the ischial rami.For other types, we used the horizontal or longitudinal centerline to measure the transverse or vertical diameter.

Image annotation
The Colabeler 2.0.4 software (KuaiYiTech, Hangzhou, China, http:// www.colab eler.com/) was used to annotate anatomical key landmarks on all pelvic radiographs (Fig. 3).The landmarks were the triradiate cartilage vertex, symphysis pubis superolateral edge, obturator margin, and lower ends of the ischial and pubic rami (if applicable).The specific annotation approach was described below.We invited nine clinicians to form a clinical team to learn uniformly about the location of landmarks and the criteria for quality assessment of pelvic radiographs.In total, 2,885 radiographs were randomly allocated to four intermediate pediatric orthopedists for annotation.Next, the expert committee, consisting of two senior pediatric orthopedists and two senior pediatric radiologists, reviewed the annotated images.The annotated radiographs were unqualified if the committee thought any key points were mislocated.In controversial cases, we consulted another senior pediatric radiologist specializing in pelvic radiographs.Unqualified radiographs were re-annotated until they met the experts' requirements.

Network framework
The proposed AI model, named the 'SN-APR', adopted the standard Mask R-CNN framework to detect anatomical key landmarks and segment the obturator region (Fig. 4).We converted the detection of an anatomical landmark to a task of identifying a neighborhood area centered at the landmark.For an input pelvic radiograph (which was first preprocessed), the ResNet-50-FPN backbone was employed to extract a feature map.A region proposal network generated bounding box proposals based on this feature map.Besides, the RoIAlign layer combined the feature map and bounding box proposals to convert the feature map of the region of interest into a feature map with certain dimensions.The generated feature maps were then fed into the head network (including the classifier branch and mask branch).The 7 × 7 256-D feature maps were put into the classifier branch to predict class labels for the bounding boxes and their offsets by one 7 × 7 1024-D conv layer and one 1 × 1 1024-D conv layer.The bounding-box offsets were decoded to obtain the location information of the bounding boxes, identifying specific anatomical landmarks at the center of corresponding boundingboxes.Meanwhile, the 14 × 14 256-D feature maps were put into the mask branch, consisting of a stack of four 3 × 3 256-D conv layers, a 28 × 28 256-D deconv layer, and a 28 × 28 5-D conv layer, to classify the categories of each pixel and segment the obturator foramen region.Apart from the background, there were nine classifications (num_classes = 9): the area near each of the seven anatomical landmarks and the region of bilateral obturator foramen were regarded as separate categories.

Performance test
The evaluation results of the testing set by clinicians were used as a standard to evaluate the diagnostic performance of the model.We used the image analysis software Digimizer 5.7.2 (MedCalc Software, Ostend, Belgium) to simulate the conventional quality assessment of pelvic radiographs.The main steps were as follows: 1) marking the top of bilateral triradiate cartilage and the superolateral edge of the symphysis pubis, 2) drawing the reference line (e.g., Hilgenreiner's line), 3) estimating obturator transverse and vertical diameters, 4) calculating the PRI and PTI, and 5) making final judgments.A pelvic radiograph was standard only if the PRI and PTI fit within the reference range of 0.56-1.80 and 0.75-1.20;otherwise, it was a non-standard radiograph.Moreover, we compared the difference in time consumption between the AI model and clinicians.

Statistical analysis
This study measured the PTI, PRI, and final judgment (i.e., standard or nonstandard radiographs) to compare the performance difference between the AI model and clinicians.The receiver operating characteristic (ROC) curves were conducted to evaluate the model's performance in the quality assessment of pelvic radiographs.Furthermore, the Bland-Altman plots and independent t-tests were applied to assess the agreement of PTI and PRI measurements between the AI model and clinicians.Statistical significance was set at p < 0.05.All data were statistically analyzed using SPSS 28.0 (IBM, Armonk, United States) and GraphPad Prism 9.3.1 (GraphPad, San Diego, United States).

Training and modification
This study used 2,885 anteroposterior pelvic radiographs to train and validate the AI model.Utilizing the PyTorch framework, we modified the open-source project Mask-R-CNN (https:// github.com/ pytor ch/ vision/ blob/ main/ torch vision/ models/ detec tion/ mask_ rcnn.py) to fit the training targets of the present study.
The training was initiated using the pre-trained weights based on the COCO 2017 dataset.We use stochastic gradient descent with an initial learning rate of 0.004 for 100 epochs and 0.001 for the next 200 epochs on the dataset to train the SN-PXR.The momentum and weight decay are set to 0.9 and 0.0001, respectively.All training and testing for the SN-PXR are performed In the modification phase, we found that the measured diameters in some misdiagnosed cases did not intersect with the obturator edges.The principal reason was an error in calculating obturator diameters, which failed to consider the influence of obturator defects in young children sufficiently.Accordingly, the calculation technique of obturator diameters was refined by actively pre-avoiding the intersection with obturator defects.The specific steps are as follows: extracting the obturator contours, deleting the contours of the obturator defects according to the coordinates of the obturator lower ends, obtaining the minimum outer rectangle of the right obturator foramen, and walking through the points on the lateral side of the rectangle in turn.

Model effect evaluation
The testing set included 362 cases (female: male = 208:154; mean age 3.93 ± 2.61 years).These cases were used to calculate the PTI and PRI, identify possible pelvic tilt and rotation, and evaluate the quality of pelvic radiographs.Comparative assessments of the quality of pelvic radiographs by the AI model and clinicians are detailed in Table 1.According to Tönnis' and Ball's criteria, the entire dataset was classified as the standard group (absence of pelvic malposition) and the non-standard group with pelvic tilt or rotation.The specific results of the quality assessment are shown in Table 2.There were two cases evaluated as 'standard radiographs' by clinicians but 'non-standard radiographs' by the AI model (Table 3).  5. Using the undisputed diagnostic results of clinicians as the standard, the accuracy, sensitivity, specificity, and area under the subject operating characteristic curve (AUC) of our model were 99.4% (360/362), 98.6% (138/140), 100.0%(222/222), and 0.993, respectively (Fig. 6).Comparative results of PTI and PRI measurements by the AI model and clinicians are shown in Fig. 7.As for PTI measurements (362 cases), the 95% limits of agreement (Bland-Altman analyses) were -0.052-0.072(bias 0.010, p = 0.633) compared with clinical experts.As for PRI measurements, the 95% limits of agreement were -0.088-0.055(bias -0.016, p = 0.333).

Time consumption
In this study, the time consumption of assessing the pelvic radiograph included loading the image data into the proposed model or Digimizer software, plotting reference lines and points, measuring the values of PTI and PRI, and obtaining quality evaluation results.For each case in the testing set, the mean time consumption of the AI model (3.79 ± 0.15 s) was significantly less (p < 0.001) than that of the pediatric orthopedists (113.96± 3.91 s).

Discussion
Early diagnosis and timely treatment of DDH often have a good prognosis, with the prospect of restoring to a normal or near-normal hip [14].Anteroposterior pelvic radiography was one of the most common methods to secure the early diagnosis of DDH in children after the appearance of the femoral ossific nucleus [15].Reliable radiological assessment was deeply affected by pelvic malposition in threedimensional planes because of the nature of the projected image [8,9].Therefore, detecting the potential malposition from pelvic radiographs was a prerequisite for improving the clinical diagnosis of pediatric DDH.
Recent AI studies on pediatric DDH primarily focused on measurements of hip parameters and judgment of the severity of hip dislocation [16][17][18].However, they may neglect the influence of nonstandard radiographs with pelvic malposition on diagnosing DDH [8].In fact, standard pelvic radiographs, i.e., ensuring radiographs' quality, are Fig. 6 The graph shows a receiver operating characteristic (ROC) curve comparing the model's performance in screening standard pelvic radiographs versus the clinicians' performance Fig. 7 The graphs show Bland-Altman plots for A PTI (pelvic tilt index) and B PRI (pelvic rotation index) between the AI model and clinicians a prerequisite for accurate diagnosis of DDH in children.To our best knowledge, the present study is the first attempt to apply AI algorithms for screening standard pelvic radiographs and quantitative assessment of their quality.
The proposed model was highly consistent and more efficient than the conventional manual method in assessing the quality of pelvic radiographs.This technique accurately screened standard pelvic radiographs in 360 out of 362 cases (99.4%), with the sensitivity and specificity of 138/140 (98.6%) and 222/222 (100.0%),respectively.Two cases were misjudged as 'nonstandard radiographs' by the AI model due to underestimated PTI.One was challenging to locate the right triradiate cartilage vertex owing to the striking overlap of the ischium.The other, with the irregularshaped obturator inferior margin, made the precise calculation of Dv difficult.However, when the original radiographs were appropriately cropped, the correct quality evaluation and accurate parametric measurements were made (Supplementary Fig. 1).We speculate that the cropped images from the original radiographs have contributed to extracting local features for the AI model.Besides, there may be more images with a cropped image-like size in the training dataset.Thus, it is recommended not to scale images but to crop out their unnecessary parts for some tricky cases.
In the testing set, the incidence of non-standard radiographs was approximately 61.3% (222/362), including 218 cases with pelvic tilt (PTI 0.670 ± 0.328), 1 case with pelvic rotation (PRI 2.108), and 3 cases with combinations of pelvic tilt and rotation (PTI 0.624 ± 0.040, PRI 1.753 ± 1.160).We thought anteroposterior pelvic tilt occurred more subtly during radiography due to the children not lying flat or to the difference in thickness of their clothes.This finding reminds clinicians to pay close attention to the potential pelvic tilt while performing pelvic radiography or diagnosing hip disorders.
In this study, the quality of pelvic radiographs was determined by investigating possible pelvic rotation and anteroposterior tilt.We used Ball's pelvic tilt index (PTI) [13] rather than Tönnis' symphyseal-ischial angle (SIA) [19] to assess anteroposterior pelvic tilt in the sagittal plane.The SIA was formed by two tangent lines from the highest point of each ischium to the symphysis point.The sciatic highest point was easily obscured by the superior ramus of the pubis due to the nature of the two-dimensional projected image.Our pre-experiment results showed significant differences in the localization of this point between observers, leading to measurement errors in calculating the SIA.Ozyalvac et al. adopted the angle between two lines connecting each symphysis of the upper arm (not the sciatic highest point) and the symphysis point to calculate the SIA [20].In addition, as a strictly age-related index, the normal values of the SIA range from 85° to 135°.In fact, it was tough to accurately predict a child's physical age from the pelvic radiograph, even for experienced pediatric orthopedists.Besides, the SIA applies only to children under five years, while the PTI used in this study is more widely applicable.Meanwhile, the PTI and PRI share several anatomical landmarks (e.g., the triradiate cartilage vertex, obturator margin, and Hilgenreiner's line), which makes the calculating process more efficient and convenient.Therefore, it is not feasible for the AI model to use the SIA to determine the quality of pelvic radiographs.
When first proposed by Tönnis, the PRI was not precisely delineated beyond its definition [19].Prudencia et al. further clarified that bilateral transverse diameters are the maximum horizontal width of the obturator foramina [21].However, this condition may be difficult to meet due to potential pelvic inclination in the coronal plane during pelvic radiography.The pelvic inclination can make the horizontal line of a radiograph unparallel to the Hilgenreiner's line.Blindly relying on the horizontal line to determine obturator vertical and transverse diameters would greatly reduce the effectiveness of the measurements.Thus, we suggested using the Hilgenreiner's line as a reference to measure the obturator transverse and vertical diameters.
We do realize some limitations of the current study.First, the criteria of quality assessment were affected by the obturator morphology and closure of the triradiate cartilage.Our study may not be generalized to children with premature closure of the triradiate cartilage.It is generally accepted that older children are more cooperative during pelvic radiography.Secondly, inaccurate measurements may occur in cases with blurry landmarks due to extreme pelvic malposition.Besides, our model was unfit for patients after pelvic osteotomy due to internal fixations and potential displacement of the triradiate cartilage.Lastly, the dataset size of the study was relatively small.The AI model needs further learning and improvement in parametric measurements.

Conclusion
In summary, the proposed model was the first attempt to apply AI algorithms to screen standard anteroposterior pelvic radiographs with great consistency and efficiency.This method can be used for real-time validation of the quality of pelvic radiographs in current PACS, which will remarkably improve the accuracy of interpreting pelvic radiographs and diagnosing DDH.The initial success of our automatic assessment method lays the groundwork for developing fullintelligent comprehensive diagnostic techniques for pediatric hip disorders.

Fig. 1
Fig.1The flowchart shows inclusion/exclusion criteria and three datasets' distribution in the study.AI, artificial intelligence

Fig. 3 Fig. 4
Fig. 3 This image shows obturator morphology for young children at different stages of skeleton calcification.The lower end of the ischium (I-point, black point); the center obturator vertical line (COVL, white dotted line)

Fig. 5
Fig. 5 These images display examples of quality assessment of pelvic radiographs for clinicians and the AI model in A obturator type 2 (female, 1.6 years), B obturator type 3 (male, 1.6 years), and C obturator type 4 (male, 3.8 years).The left images represent the evaluation results of clinicians, while the right ones represent those of the AI model

Table 1
Patient demographics and results of quality assessment

Table 2
Distribution of results for quality assessment of pelvic radiographs

Table 3
Two cases misdiagnosed as 'nonstandard radiographs' by the AI model a Values for calculating pelvic tilt and rotation index (PTI and PRI) are in pixels on a computer with NVIDIA GeForce GTX 1660 GPU SUPER and Intel i7-10700 CPU (Santa Clara, California, United States).