Article Text
PDF
Abstract
Introduction Radiographic bone age (BA) assessment is widely used to evaluate children’s growth disorders and predict their future height. Moreover, children are more sensitive and vulnerable to X-ray radiation exposure than adults. The purpose of this study is to develop a new, safer, radiation-free BA assessment method for children by using three-dimensional ultrasound (3D-US) and artificial intelligence (AI), and to test the diagnostic accuracy and reliability of this method.
Methods and analysis This is a prospective, observational study. All participants will be recruited through Paediatric Growth and Development Clinic. All participants will receive left hand 3D-US and X-ray examination at the Shanghai Sixth People’s Hospital on the same day, all images will be recorded. These image related data will be collected and randomly divided into training set (80% of all) and test set (20% of all). The training set will be used to establish a cascade network of 3D-US skeletal image segmentation and BA prediction model to achieve end-to-end prediction of image to BA. The test set will be used to evaluate the accuracy of AI BA model of 3D-US. We have developed a new ultrasonic scanning device, which can be proposed to automatic 3D-US scanning of hands. AI algorithms, such as convolutional neural network, will be used to identify and segment the skeletal structures in the hand 3D-US images. We will achieve automatic segmentation of hand skeletal 3D-US images, establish BA prediction model of 3D-US, and test the accuracy of the prediction model.
Ethics and dissemination The Ethics Committee of Shanghai Sixth People’s Hospital approved this study. The approval number is 2022-019. A written informed consent will be obtained from their parent or guardian of each participant. Final results will be published in peer-reviewed journals and presented at national and international conferences.
Trial registration number ChiCTR2200057236.
- Diagnostic Imaging
- PAEDIATRICS
- Child protection
- Paediatric radiology
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
STRENGTHS AND LIMITATIONS OF THIS STUDY
This is the first prospective study to build up a non-ionising radiation three-dimensional (3D) ultrasound bone age prediction model by artificial intelligence (AI) and to examine its diagnostic accuracy.
This study combines knowledge from four different disciplines and novel technical methods, such as a newly developed 3D-US scanning equipment and fully automatic 3D image recognition and bone age analysis technology based on AI.
Statistician will be blinded to the outcome measures and relevant parameters of all participants, the results will be objective.
This study currently only includes a single centre.
Introduction
The phenomenon of precocious puberty in contemporary Chinese children is becoming increasingly common and has aroused widespread concern. Precocious puberty affects children’s stature and physical development, seriously endangering children’s healthy growth. Bone age (BA) has been accepted as an important criterion for assessing growth and development and diagnosing sexual precocity in children. Radiographic BA assessment is often used to evaluate growth and puberty in children and adolescents as part of the routine diagnostic work.1 2 Radiography is radioactive, children are more sensitive to the ionising radiation of X-rays than adults, and the younger they are, the more sensitive they are, especially girls.3 Repeated BA assessments are usually performed during the follow-up of children with growth and puberty problems. Low-dose X-ray examination is often used for BA assessment at present, but there are still potential risks of radiation.4 The cumulative effects of ionising radiation on children have caused great anxiety among parents and paediatricians. Unfortunately, it remains a difficult problem to find a safer and more reliable method for BA assessment.
Over the past three decades, many researchers have explored the application of ultrasound (US) to estimate BA in order to address the potential hazard caused by repeated irradiation.5–8 Some researchers have used portable automatic ultrasonic devices to assess BA, such as commercially available BAUS (SonicBone, Rishon Lezion, Israel) and BonAge system (Sunlight Medical, Tel Aviv, Israel).1 9 However, these automatic ultrasonic devices only measured velocity changes of ultrasonic waves through the epiphysis of ulna, radius, proximal third phalanx, and metacarpal bones. Due to the lack of data acquisition, the accuracy of BA assessment is obviously affected and these systems cannot even display US image.5 6 10 11 In addition, conventional US machines have also been used to analyse BA by other researchers. Conventional US can clearly demonstrate the hyperechoic ossification centres and the hypoechoic epiphysial plate cartilage which cannot be detected by radiography.5 12 Based on the reproducible appearance of a growing epiphysial plate and characteristic echogenicity of ossification centre, Bilgili et al5 presented the US version of the Greulich-Pyle atlas for children with definitely known precocious or delayed skeletal maturation. Hajalioghli et al adopted a similar method to detect BA abnormalities in children.12 Wan et al measured ossification ratios of 13 bones of the hand and wrist in newborn to near-adult subjects by using conventional US.13 The skeletal maturity scoring system which derived from the ossification ratios of the 13 bones may provide a quantitative modality to evaluate BA. These studies demonstrated that the radiographic BA assessment can be replaced by US. However, none of these methods can provide intuitive high-quality visual images for clinical paediatricians and endocrinologists to observe and analyse. Meanwhile, the entire process of image and data acquisition is very time-consuming in these studies. A good method for estimating BA should be accurate, objective, convenient, and time-saving.14 15
To address the limitations of two-dimensional (2D) ultrasonic BA assessment in previous studies, we hypothesised that three-dimensional ultrasound (3D-US) could be applied to BA assessment. 3D-US is a powerful imaging technique that can provide a more complete and detailed view of anatomical structures than 2D-US,16 which only shows partial views of the complex 3D.17 18 3D-US is widely used in prenatal US, echocardiography, automated breast US, musculoskeletal US, and other medical fields.17 18 Compared with conventional 2D-US, 3D-US has the following advantages: accurate positioning and high precision,19 20 reducing the dependence on the technical level of the operator and enhancing the repeatability, providing more anatomical structure details and relevant parameters, and improving imaging quality, measurement, and diagnostic level. 3D-US can not only provide clinicians with clear and intuitive 3D images but also obtain more details and data for subsequent computer processing.
Further, in order to overcome the inefficiency, inexperience, and skill deficiency of some US examiners,21 22 we propose to apply artificial intelligence (AI) to skeletal identification and segmentation in 3D-US skeletal images, and to estimate BA through the relevant parameters of different epiphyses in the hand. The medical application of AI has gained considerable attention and made great breakthroughs over the past few years.23 There have been many reports on the use of AI in radiological research, and AI-assisted diagnosis has gradually become a research hotspot in the field of US.24–27 Some experts have gained great progress in the AI-assisted US diagnosis of liver, breast, and thyroid diseases.28–31 AI-assisted US diagnosis can significantly improve work efficiency and diagnostic accuracy through quantitative assessments, standard measurements, and automated image quality control.32 AI can shorten the acquisition time of parameters, avoid the loss of the details of image information that cannot be processed by human eyes, and improve the accuracy of data acquisition.33 34 Additionally, AI has strong automatic recognition ability, and AI based on convolutional neural network has exemplary performance in image segmentation and reconstruction.35–37 It is widely reported that the combination of AI neural networks and 3D breast US has achieved great success in the diagnosis of breast diseases.38–40 It can also quickly, accurately, and automatically accomplish the segmentation and reconstruction of radiographic images of BA.41–43 To our best knowledge, however, no study has used AI-assisted the skeletal 3D-US in the assessment of BA.
Therefore, this study is designed to obtain left-hand perspective skeletal images of children of different ages by 3D-US automatic scanning, achieve automatic segmentation of skeletal 3D-US images, establish a regression model for BA prediction of 3D-US, and test the accuracy of the prediction model. This study provides a new, safer and more efficient method for BA assessment in children by applying AI-assisted 3D-US diagnosis and test the diagnostic accuracy of this method.
Methods and analysis
Study design
This is a prospective, observational study that will enrol participants who meet the inclusion criteria in the Paediatric Growth and Development Clinic. Participants will undergo BA assessment based on hand 3D-US scanning equipment and AI recognition in Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. Using radiographic BA as the reference standard, consistency evaluation will be used to test the accuracy of ultrasonic BA. This study follows Standards for the Reporting of Diagnostic Accuracy Studies guidelines.44
Participant and public involvement
Participants were not involved in the design, or conduct, or reporting, or dissemination plans of our research. The final results and related publications will be widely disseminated to the public through mass media related to this field. Participants will be acknowledged at the end of our publications and presentations.
Participant recruitment
Participants will be recruited from patients who need X-ray BA assessment due to various personal reasons at the Paediatric Growth and Development Clinic of Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. The participant flow chart of this study is shown in figure 1.
Participants recruitment and participants flow chart. 3D-US, three-dimensional ultrasonography; AI, artificial intelligence; BA, bone age.
Medical evaluation and enrolment procedure
Inclusion criteria
Age in the range of 6–18 years old;
No pathology in the hand and wrist (no acute trauma, fracture, arthritis, and deformity).
Complete the 3D-US scanning and radiography, and their parents sign the written informed consent.
Exclusion criteria
Having history of pathology in the hand and wrist.
Having history of hermaphroditism.
Having mental retardation and cannot cooperate.
Demographic variables regarding age, gender, ethnicity, and health condition will be recorded before US scanning of all participants.
Radiographic BA
Radiography of the children’s left hand and wrist of will be performed using the same equipment (Philips, digital diagnosis C90, Philips Medical Systems Nederland B.V., Nederland) in the radiology department of Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine. During the radiographic examination, the appropriate position will be provided for participants according to the standards determined by the equipment manufacturer, the parameters are set to an average of 66 kVp, 5 mA and 6 S. The estimated BA will be interpreted with the TW3-C RUS method (third edition of Tanner-Whitehouse method-the radius, ulna, and short bones of the first, third, and fifth fingers). Radiographic BA will be estimated by two experienced paediatricians according to TW3. The average value of the radiographic BA results evaluated by two experienced paediatricians will be considered as gold standard.
Ultrasound scanning
All participants will undergo 3D-US scanning and other routine necessary tests of the Paediatric Growth and Development Clinic (including radiographic BA) on the same day. A new ultrasonic scanning device will be used to carry out automatic 3D-US scanning of hands for children of different ages. The new device adopts 2D sequence automatic scanning similar to Automated Breast Volume Scan volumetric imaging, forming 3D image without reducing spatial resolution and contrast. The frequency of the transducer in the device is 12 MHz and the width of the transducer is 133 mm. The device is a box type, with automatic probe scanning, volumetric imaging, and coupling agent is water not gel. An illustration of the device is shown in figure 2. This new automatic 3D-US scanning device has received the medical device registration certificate of the People’s Republic of China. When 3D-US scanning is finished, AI will be trained to automatically identify and mark the ossification centres and epiphyses of the hand bone in the optimised perspective 3D ultrasonic skeletal imaging. According to TW3-RUS method which associated with the evaluation of BA, relevant parameters will be measured at ossification centres and epiphyses of 13 bones in the hand and wrist.45 The statistician will not be involved in the ultrasonic examination and images analysis procedures.
Diagrammatic sketch of the new automatic 3D-US scanning device. 1, ultrasound machine; 2, container with water as coupling medium; 3, automatic scanning probe; 3D-US, three-dimensional ultrasonography; AI, artificial intelligence.
AI research process
The 3D US scanning image data of patients aged 6–18 years old will be collected and randomly divided into a 20% test set and an 80% training set.
To save computing resources, the image data will be resampled, meaning that the cross section will be resized to 512×256 while keeping the vertical axis size unchanged. Approximately 450 valid slices will be obtained for each case, totalling around 738 000 slices used for training the image segmentation model. The BA prediction model is trained with the entire case as input, resulting in a total of 1640 patches for training. To enhance the variability of the data representation and improve the generalisation of the model, random crop, random translation, and rotation will be used to enhance the data. The network will be trained and tested using PyTorch on an NVIDIA GeForce RTX 3090 GPU (NVIDIA Corp., Santa Clara, CA).
Because some artefacts and noise in the US image can affect the prediction performance, the U-Net method will first be used to segment the US image into multiple categories, identify the hand bone region, and subdivide it into different bones. Then, the gender of the sample will be encoded by one-hot encoding. The results of image segmentation and the gender coding will be fused and extracted by using the transformer model based on multihead attention. Finally, the global average pooling layer will be used to reduce the dimension of the extracted features, and the prediction results of BA will be output through the full connection layer.
The above models of image segmentation and BA prediction will be combined into a cascade network to realise the end-to-end prediction of image to BA. At the same time, the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm will be used to generate a heat map according to the gradient and activation of the feature map. This will visualise the area of concern of the model and provide interpretability for the model prediction. Dice and average symmetric surface distance (ASSD) will be used as the evaluation index of image segmentation, and mean absolute error (MAE) will be used as the evaluation index of BA prediction.
Study regimen and statistical analysis
Of the total data, 80% will be randomly extracted as the training set and 20% of the total data will be used as the test set. The study has been divided into two parts according to the objectives. The statistical analysis is presented in each of the substudies, and statistical analysis will be performed by the responsible study biometrician.
Objective 1: To build up a 3D ultrasound bone age (BA) model of AI
Training set will be involved in this substudy. X-ray and US image data of the left hand will be recorded, and radiographic BA evaluated by two experienced paediatricians will also be recorded. The average value of the radiographic BA results evaluated by two experienced paediatricians will be considered as gold standard. After preprocessing of standardised 3D US image and data enhancement, the image features, gold standard, and gender will be combined to establish BA predictive regression model of AI in State Key Laboratory of Mechanical System and Vibration (Shanghai Jiao Tong University). The specific methods and algorithms are listed in AI research process.
Objective 2: To evaluate the performance of the AI model
The test set will be involved in this substudy. First, the MAE will be used to evaluate the accuracy of BA prediction model. In addition, two other paediatricians will participate in this part of the study. The average value of radiographic BA obtained by these two paediatricians will be regarded as BA evaluated by paediatricians and used for data analysis. The statistical unit of radiographic and ultrasonic BA are years, and figures will be accurate to one decimal place in the output data. Paired t-tests will be used to compare the results of predictive BA between the two methods and the gold standard. p<0.05 is considered to indicate a statistically significant difference. Consistency evaluation will be used to assess the consistency between BA evaluated by AI model, BA evaluated by paediatricians, and the gold standard. Pearson’s correlation analysis will be used to analyse the correlation of the two methods and the gold standard. Regression analysis will be performed for the analysis of the relationship between the AI model and the gold standard, CIs for regression coefficients and intercept will be calculated. Scatter plots and Bland Altman graphs will be used to demonstrate the consistency between the three sets of data. SPSS V.24.0 and GraphPad Prism V.9.0 will be used for statistical analysis.
Descriptive statistics will be used in the following multiple variables: gender, age, height, weight, and BA. The data analysis of the test set will be categorised according to three methods (AI model, paediatricians, and the gold standard), which will be conducted blinded to the outcome assessors and statisticians. All statisticians will be blinded to the participants’ clinical data.
Sample size calculation
Theoretically, the larger sample size used to establish an AI prediction model of 3D-US, the better the prediction effect. Considering that the sample size increases to a certain extent, the accuracy of prediction may not improve significantly. We plan to determine the sample size based on the performance of the AI model, which means that when the ultrasonic BA predicted by AI model and radiographic BA correlated well, the sample size can meet the requirements. Based on previous similar literatures of ultrasonic BA, ultrasonic BA and radiographic BA correlated well in the study of Zhao et al (Pearson’s r=0.96 and 0.96 for girls and boys),46 and in another research (Pearson’s r=0.97 and 0.97 for girls and boys).13 A target sample size of n=369 can achieve 80% power to detect a difference of −0.01000 between the null hypothesis correlation of 0.96000 and the alternative hypothesis correlation of 0.97000 using a two-sided significance level of 0.05000. With an estimated dropout rate of 10%, we aim to recruit n=410 patients in the validation group of this study. The validation group, which is considered as the test set, accounts for 20% of the total data, so we plan to enrol at least n=2050 participants in total in this study.
Data management
Demographic and somatic data will be collected before 3D ultrasonic scanning and images processing. Baseline characteristics of participants will be presented with descriptive results. The statistical unit of BA is year, and figures will be accurate to one decimal place in the output data.
Data will be stored in computer and password-protected. Access to the data will be limited to the investigators and clinical trial staff. The information and data are available at any times for inspection by the ethics commission.
Participants will be withdrawn from the research immediately if: (1) the participants withdraw their consent, and (2) exclusion criteria are found after enrolment. The reasons for discontinuation will be recorded, along with the dates of discontinuation. The withdrawn participant’s data will be unavailable in this study.
Adverse events
The 3D-US examination will not cause any reactions to participants in deed, since it is a risk-free observational study. Adverse events will be recorded if existed.
Quality assurance
All staff involved in this study have the professional knowledge and experience required in the study. Investigators and other staff of this study will follow the clinical study protocol and adopt standard operating procedures to ensure the quality control of the study. The investigators, paediatricians, research assistants, statisticians, and data assessors are different people. They should all receive Good Clinical Practice training. All obtained data and images will be stored electronically and kept strictly confidential with secured access. The participant’s personal information will be kept confidentially in accordance with the legal requirements. When necessary, the government management department, the hospital ethics committee, and its relevant personnel can consult the participant’s data according to the regulations and relevant laws.
Study duration
The study has started in September 2022, and all participants’ recruitment are expected to be completed by December 2024. Figure 3 shows study procedures.
Study design flow chart. 3D-US, three-dimensional ultrasonography; AI, artificial intelligence; BA, bone age.
Ethics and dissemination
This study was approved by Ethics Committee of Shanghai Sixth People’s Hospital. The approval number is 2022-019. This study will be conducted according to the principles of the Declaration of Helsinki. No interventional therapy, 3D-US scanning in the clinic study is non-radioactive. It is considered to be no potential risks in the study and is addressed in the protocol and written consent forms. Parents or guardians of each participant will sign a written informed consent. The personal information of the participants will not be disclosed unless authorisation is approved.
Results will be published in peer-reviewed scientific publications and presented at national and international conferences.
Limitation
This study has limitations. This study currently only includes a single centre, the sample from the region of Shanghai cannot be applicable to all Chinese. In the future, multiple research centres in different regions may participate in this research and the external validation of this study.
Ethics statements
Patient consent for publication
Acknowledgments
The authors thank the children and adolescents who will take part in the study, and their parents and families.
References
Footnotes
Contributors YYZ is the primary investigators. LC, BLZ, JS, MW, XJC, and YYZ were involved in the development of the study design. LC, JS, JCX, BLZ, ZHC, SDS, JC, XJC, TY, BH, MW, XJC, and YYZ participated in the study conduct. LC, BLZ, and JS drafted the manuscript under YYZ’s supervision. YYZ contributed to applying for and gaining funding. All authors contributed to the content and critical revision and drafting of the manuscript. LC, BLZ, and JS contributed equally. LC, BLZ, and JS are joint first authors.
Funding This study will be supported by the Key Program of the National Natural Science Foundation of China (no. 82030050 and T2394534), National Key Research and Development Program of China (no. 2023YFC2410800 and 2022YFC3400102), Key Project of the Shanghai Committee of Science and Technology in 2021 (no. 21Y21901100).
Competing interests None declared.
Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Provenance and peer review Not commissioned; externally peer reviewed.