|
审核状态: Project audit state: |
通过审核 Successful |
|
注册号: Registration number: |
ChiCTR2500108760 |
|
最近更新日期: Date of Last Refreshed on: |
2025-09-04 16:16:18 |
|
注册时间: Date of Registration: |
2025-09-04 00:00:00 |
|
注册号状态: |
预注册 |
|
Registration Status: |
Prospective registration |
|
注册题目: |
基于超声图像的儿童腹部肿物智能诊断基础模型研究 |
|
Public title: |
Research on the basic model of intelligent diagnosis of abdominal masses in children based on ultrasound images |
|
注册题目简写: |
|
|
English Acronym: |
|
|
研究课题的正式科学名称: |
基于超声图像的儿童腹部肿物智能诊断基础模型研究 |
|
Scientific title: |
Research on the basic model of intelligent diagnosis of abdominal masses in children based on ultrasound images |
|
研究课题代号(代码): Study subject ID: |
|
|
在二级注册机构或其它机构的注册号: The registration number of the Partner Registry or other register: |
|
申请注册联系人: |
周路遥 |
研究负责人: |
周路遥 |
|
Applicant: |
Luyao Zhou |
Study leader: |
Luyao Zhou |
|
申请注册联系人电话: Applicant telephone: |
+86 134 2753 9467 |
研究负责人电话: Study leader's telephone: |
+86 134 2753 9467 |
|
申请注册联系人传真 : Applicant Fax: |
研究负责人传真: Study leader's fax: |
||
|
申请注册联系人电子邮件: Applicant E-mail: |
zhouly6@mail.sysu.edu.cn |
研究负责人电子邮件: Study leader's E-mail: |
zhouly6@mail.sysu.edu.cn |
|
申请单位网址(自愿提供): Applicant website(voluntary supply): |
研究负责人网址(自愿提供): Study leader's website(voluntary supply): |
||
|
申请注册联系人通讯地址: |
深圳市福田区益田路7019号A楼2楼超声医学科 |
研究负责人通讯地址: |
深圳市福田区益田路7019号A楼2楼超声医学科 |
|
Applicant address: |
2/F, Building A, No. 7019 Yitian Road, Shenzhen,Guangdong. |
Study leader's address: |
2/F, Building A, No. 7019 Yitian Road, Shenzhen,Guangdong. |
|
申请注册联系人邮政编码: Applicant postcode: |
研究负责人邮政编码: Study leader's postcode: |
||
|
申请人所在单位: |
深圳市儿童医院 |
||
|
Applicant's institution: |
Shenzhen Children's Hospital |
||
|
研究负责人所在单位: |
深圳市儿童医院 |
||
|
Affiliation of the Leader: |
Shenzhen Children's Hospital |
||
|
是否获伦理委员会批准: |
是/Yes |
||
|
Approved by ethic committee: |
Yes |
||
|
伦理委员会批件文号: Approved No. of ethic committee: |
深儿医伦审(科研)批件 202507502 号 |
伦理委员会批件附件: Approved file of Ethical Committee: |
查看附件View |
|
批准本研究的伦理委员会名称: |
深圳市儿童医院医学伦理委员会 |
||
|
Name of the ethic committee: |
Medical Ethics Committee of Shenzhen Children's Hospital |
||
|
伦理委员会批准日期: Date of approved by ethic committee: |
2025-08-29 00:00:00 |
||
|
伦理委员会联系人: |
李晨曦 |
||
|
Contact Name of the ethic committee: |
Chenxi Li |
||
|
伦理委员会联系地址: |
深圳市福田区益田路7019号B楼2楼深圳市儿童医院医学伦理委员会办公室 |
||
|
Contact Address of the ethic committee: |
2/F, Building B, No. 7019 Yitian Road, Shenzhen, Guangdong, Medical Ethics Committee Office |
||
|
伦理委员会联系人电话: Contact phone of the ethic committee: |
+86 755 8300 8379 |
伦理委员会联系人邮箱: Contact email of the ethic committee: |
|
|
研究实施负责(组长)单位: |
深圳市儿童医院 |
||||||||||||||||||||||
|
Primary sponsor: |
Shenzhen Children's Hospital |
||||||||||||||||||||||
|
研究实施负责(组长)单位地址: |
深圳市福田区益田路7019号A楼2楼超声医学科 |
||||||||||||||||||||||
|
Primary sponsor's address: |
2/F, Building A, No. 7019 Yitian Road, Shenzhen,Guangdong. |
||||||||||||||||||||||
|
试验主办单位(项目批准或申办者): Secondary sponsor: |
|
||||||||||||||||||||||
|
经费或物资来源: |
自筹 |
||||||||||||||||||||||
|
Source(s) of funding: |
self-financing |
||||||||||||||||||||||
|
Target disease: |
Abdominal mass |
||||||||||||||||||||||
|
Target disease code: |
|
||||||||||||||||||||||
|
研究类型: |
诊断试验 |
||||||||||||||||||||||
|
Study type: |
Diagnostic test |
||||||||||||||||||||||
|
研究所处阶段: |
其它 | ||||||||||||||||||||||
|
Study phase: |
N/A |
||||||||||||||||||||||
|
研究设计: |
诊断试验诊断准确性 |
||||||||||||||||||||||
|
Study design: |
Diagnostic test for accuracy |
||||||||||||||||||||||
|
研究目的: |
基于多模态超声数据(灰阶超声图像、彩色超声图像及临床数据)开发和验证一种能够识别儿童腹部肿物并且对儿童腹部肿物进行准确分类诊断的人工智能基础模型,并在临床真实环境中测试其帮助不同经验的超声医生提高诊断准确率的效能,探讨该模型在辅助超声医生诊断儿童腹部肿物的临床价值。 |
||||||||||||||||||||||
|
Objectives of Study: |
Based on multimodal ultrasound data (gray-scale ultrasound images, color ultrasound images and clinical data), develop and validate an artificial intelligence basic model capable of identifying abdominal masses in children and accurately classifying and diagnosing them, and test its efficacy in helping ultrasound doctors with different experiences improve diagnostic accuracy in a real clinical environment. To explore the clinical value of this model in assisting ultrasound doctors in diagnosing abdominal masses in children. |
||||||||||||||||||||||
|
药物成份或治疗方案详述: |
研究方法 1.研究现场及研究人群 研究现场:医院超声科 回顾性研究纳入截止时间:2013年1月-2025年06月 前瞻性研究时间:2025年7月-2026年7月 人群:病例组:超声发现有儿童腹部占位性病变并获得最终确定诊断的儿童/对照组:儿童腹部超声检查阴性 2. 研究设计 诊断研究/回顾性+前瞻性研究 3. 研究所需样本量 目前在基于医学图像的人工智能研究领域中,尚未有可以参照的样本量估计方法。因此,本研究未能精确计算所需样本量。预计各中心2025年12月前提供回顾性病例200-300例,2025年7月-2026年7月提供前瞻性病例50-100例,估计共收集回顾性儿童腹部肿物病例2000例,前瞻性病例500-600例。其中,估计肠套叠、肾囊肿、卵巢囊肿各300-400例,肝脏局灶性结节增生、肝脓肿、腹腔脓肿各150例,肝母细胞瘤、肾母细胞瘤、腹膜后神经母细胞瘤、肝血管瘤、肝间叶错构瘤、脾脏血管瘤、肠重复、卵巢畸胎瘤各100例,余相对罕见病例各50例。 4. 资料收集 收集入组儿童年龄、性别、体重(g)、血清肿瘤标记物、血常规(血红蛋白、白细胞、C反应蛋白)、肝功能、肾功能、超声图像和视频、CT或MRI诊断结果、随访结果、手术发现、病理诊断结果。 超声扫描仪的品牌和版本以及探头的类型和频率也需要记录。制定统一的前瞻性病例超声图像和视频采集标准操作流程,统一培训并定期反馈,确保多中心执行一致性。 前瞻性病例超声图像和视频采集 每例均需存2套 线阵探头+凸阵探头 图像要求:扇形探头:深度>肿物深径,增益适中,尽量显示肿物全貌 高频探头(肿物过大时,高频探头无需显示肿物全貌):频率>7MHz,深度≤5cm,增益适中。 彩超:scale<6cm/s,避免血流外溢或伪彩,可采用SMI模式。 每套图包含4-5幅静态图+2个动态视频 灰阶图像3-4张:肿物最大切面1张+与最大切面正交垂直切面图1张+肿物与周围脏器交界图1-2张 彩超图像2张:与灰阶图像对应的彩色血流图 动态视频:灰阶图像动态视频1个+彩超动态视频1个,尽量包含整个肿物体积,视频时长6s左右 如有多个肿瘤,多个肿瘤均应按照上述方式进行彩图。 所有图像和视频都应保存为DICOM格式。 回顾性病例图像采集 由一名经验>5年的资深超声医生对所有符合条件的儿童腹部肿物超声图像进行审查,在排除低质量图像后,收集所有与腹部肿物相关的图像进行进一步分析。 5.数据管理与统计分析方案 数据上传与管理:采用坚果云云盘提交数据,坚果云盘专人管理。临床一般资料,人口学信息,生化指标录入Excel电子表格;建立云盘统一上传和管理多中心超声图像数据;采用高年资超声专家读片的方式对不符合要求的超声图像进行清洗。 统计分析:①采用t检验,Chi-square检验,Delong检验,加权kappa检验等方法比较组间差异和观察者之间的一致性。②计算人工智能算法和不同经验超声医生诊断儿童腹部肿物良恶性和具体分类的灵敏度、特异度、阳性预测值、阴性预测值、准确率、ROC曲线下面积等。③在临床实践中,对不同经验的超声医生在无人工智能模型辅助和有人工智能辅助的条件下的诊断效能进行分析和比较,并分析人工智能模型带来的潜在获益和风险。 7.偏倚的控制 在病例入选阶段可能出现选择偏倚,采用多中心分别收集的方法,可能可以减轻选择偏倚;此外,可能遇到阴性病人和阳性病人不匹配的问题,可以在图像预处理阶段对阳性图像进行图像扩增来减少可能的偏倚。另外,也将采用盲法来减少超声医生读图时因为信息偏倚导致诊断效能偏高。 8.深度学习模型 本研究的深度学习模型建立和验证将由蓝网公司负责。数据提交给蓝网公司前,首先进行去标签化处理,消除病人姓名、住院号等可以识别个人身份的信息后再交给蓝网公司处理。并提前与蓝网公司签署合作协议,确保数据仅用于本次研究。 将采用ImageNet,ResNet,deeplab,YOLOv8,Mask R-CNN,U-Net, transformer等图像识别和分割模型作为骨架进行人工智能深度学习模型的构建,并根据预训练结果进行微调。深度学习模型将采用如下方法逐步构建:首先基于临床数据和多模态超声图像训练识别多个儿童腹部肿物和鉴别腹部肿物良恶性的模型,选取外部验证集上性能最佳的作为最终模型。在鉴别良恶性的基础上,将良性肿物分为囊肿、炎症病变和良性肿瘤性病变,将恶性肿物根据病理类型进行初步分类(肝母细胞瘤、肾母细胞瘤、神经母细胞瘤,等等)。并采用图像翻转等方式对阳性病例图像进行过采样。采用集成深度学习来提升模型的延展性。 将多个中心收集的样本随机分为训练集和数个验证子集,用于训练和验证模型的诊断效能。从另外几个中心(这些中心的样本从未参与训练与验证)收集样本作为外部验证集,随机分为外部验证集1和外部验证集2,进一步验证模型的延展性。 |
||||||||||||||||||||||
|
Description for medicine or protocol of treatment in detail: |
Research Methods 1. Research Site and Population Research Site: Ultrasound Department of the Hospital Retrospective study inclusion period: January 2013 - June 2025 Prospective study period: July 2025 - July 2026 Population: Case group: Children with abdominal masses detected by ultrasound and with a final confirmed diagnosis / Control group: Children with negative abdominal ultrasound results 2. Research Design Diagnostic study / Retrospective + Prospective study 3. Required Sample Size At present, there is no sample size estimation method available for reference in the field of artificial intelligence based on medical images. Therefore, the required sample size for this study could not be precisely calculated. It is expected that each center will provide 200-300 retrospective cases by December 2025 and 50-100 prospective cases from July 2025 to July 2026. It is estimated that a total of 2,000 retrospective cases of abdominal masses in children and 500-600 prospective cases will be collected. Among them, it is estimated that there will be 300-400 cases each of intussusception, renal cysts, and ovarian cysts, 150 cases each of focal nodular hyperplasia of the liver, liver abscess, and abdominal abscess, and 100 cases each of hepatoblastoma, nephroblastoma, retroperitoneal neuroblastoma, hepatic hemangioma, hepatic mesenchymal hamartoma, splenic hemangioma, intestinal duplication, and ovarian teratoma. The remaining relatively rare cases will be 50 cases each. 4. Data Collection Collect the age, gender, weight (g), serum tumor markers, blood routine (hemoglobin, white blood cells, C-reactive protein), liver function, kidney function, ultrasound images and videos, CT or MRI diagnosis results, follow-up results, surgical findings, and pathological diagnosis results of the enrolled children. The brand and version of the ultrasound scanner and the type and frequency of the probe also need to be recorded. Develop a unified standard operating procedure for the collection of prospective case ultrasound images and videos, provide unified training, and provide regular feedback to ensure consistency in multi-center implementation. Prospective case ultrasound image and video collection Each case requires 2 sets: linear probe + convex probe Image requirements: Sector probe: depth > depth of the mass, moderate gain, try to display the entire mass High-frequency probe (when the mass is too large, the high-frequency probe does not need to display the entire mass): frequency > 7 MHz, depth ≤ 5 cm, moderate gain. Color Doppler: scale < 6 cm/s, avoid blood flow overflow or pseudo-color, SMI mode can be used. Each set of images should contain 4-5 static images + 2 dynamic videos Gray-scale images 3-4: 1 image of the largest section of the mass + 1 image of the section perpendicular to the largest section + 1-2 images of the junction of the mass and surrounding organs Color Doppler images 2: color blood flow images corresponding to the gray-scale images Dynamic videos: 1 gray-scale dynamic video + 1 color Doppler dynamic video, try to include the entire mass volume, video duration about 6 seconds If there are multiple tumors, all tumors should be imaged in color according to the above method. All images and videos should be saved in DICOM format. Retrospective case image collection All ultrasound images of abdominal masses in eligible children were reviewed by a senior ultrasound doctor with more than 5 years of experience. After excluding low-quality images, all images related to abdominal masses were collected for further analysis. 5. Data Management and Statistical Analysis Plan Data upload and management: Data were submitted via Nuts Cloud Disk and managed by a dedicated person. General clinical data, demographic information, and biochemical indicators were entered into Excel spreadsheets. A unified cloud disk was established for uploading and managing multi-center ultrasound image data. Unqualified ultrasound images were cleaned by high-experience ultrasound experts. Statistical analysis: ① t-test, Chi-square test, Delong test, weighted kappa test, and other methods were used to compare differences between groups and consistency among observers. ② Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and area under the ROC curve for the diagnosis of benign and malignant abdominal masses in children by artificial intelligence algorithms and ultrasound doctors with different levels of experience were calculated. ③ In clinical practice, the diagnostic efficacy of ultrasound doctors with different levels of experience under conditions with and without artificial intelligence model assistance was analyzed and compared, and the potential benefits and risks brought by the artificial intelligence model were analyzed. 7. Bias Control Selection bias may occur during the case selection stage. A multi-center approach for data collection may help mitigate selection bias. Additionally, there may be a mismatch between negative and positive cases. To address this, image augmentation can be applied to positive images during the image preprocessing stage to reduce potential bias. Moreover, a blinded approach will be adopted to minimize the overestimation of diagnostic performance due to information bias when ultrasound doctors interpret the images. 8. Deep Learning Model The establishment and validation of the deep learning model in this study will be handled by BlueNet Company. Before submitting the data to BlueNet Company, de-labeling will be performed to remove personal identification information such as patient names and hospital numbers. A cooperation agreement will be signed in advance with BlueNet Company to ensure that the data is used only for this study. Image recognition and segmentation models such as ImageNet, ResNet, deeplab, YOLOv8, Mask R-CNN, U-Net, and transformer will be used as the backbone for constructing the artificial intelligence deep learning model, and fine-tuning will be conducted based on the pre-training results. The deep learning model will be constructed step by step as follows: First, multiple models for identifying abdominal masses in children and differentiating benign and malignant abdominal masses will be trained based on clinical data and multimodal ultrasound images. The model with the best performance on the external validation set will be selected as the final model. On the basis of differentiating benign and malignant masses, benign masses will be classified into cysts, inflammatory lesions, and benign neoplastic lesions, and malignant masses will be preliminarily classified according to pathological types (hepatoblastoma, nephroblastoma, neuroblastoma, etc.). Positive case images will be oversampled using methods such as image flipping. Ensemble deep learning will be employed to enhance the model's generalizability. Samples collected from multiple centers will be randomly divided into a training set and several validation subsets for training and validating the model's diagnostic performance. Samples from several other centers (which have never participated in training or validation) will be collected as external validation sets and randomly divided into external validation set 1 and external validation set 2 to further verify the model's generalizability. |
||||||||||||||||||||||
|
纳入标准: |
1)0-18岁儿童; 2)超声发现有儿童腹部占位性病变; 3) 腹部占位性病变通过随访、手术、活检或临床综合信息获得确诊。 |
||||||||||||||||||||||
|
Inclusion criteria |
1) Children aged 0 to 18; 2) Ultrasound detected space-occupying lesions in the child's abdomen; 3) Abdominal space-occupying lesions are diagnosed through follow-up, surgery, biopsy or comprehensive clinical information. |
||||||||||||||||||||||
|
排除标准: |
1)最终诊断不明确; 2) 超声图像质量差 |
||||||||||||||||||||||
|
Exclusion criteria: |
1) The final diagnosis is unclear; 2) Poor quality of ultrasound images |
||||||||||||||||||||||
|
研究实施时间: Study execute time: |
从 From 2025-08-29 00:00:00至 To 2026-08-29 00:00:00 |
征募观察对象时间: Recruiting time: |
从From 2025-09-05 00:00:00 至 To 2026-08-29 00:00:00 |
|
诊断试验: Diagnostic Tests: |
|
||||||||||||||||||||||||||||
|
研究实施地点: Countries of recruitment and research settings: |
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
测量指标: Outcomes: |
|
|
采集人体标本:
Collecting sample(s)
|
|
|
征募研究对象情况: Recruiting status: |
尚未开始 Not yet recruiting |
年龄范围: Participant age: |
|
||||||
|
性别: |
男女均可 |
Gender: |
Both |
||||||
|
随机方法(请说明由何人用什么方法产生随机序列): |
无 |
||||||||
|
Randomization Procedure (please state who generates the random number sequence and by what method): |
N/A |
||||||||
|
是否公开试验完成后的统计结果: Calculated Results after the Study Completed public access: |
公开/Public |
|
盲法: |
无 |
|
Blinding: |
None |
|
试验完成后的统计结果(上传文件): |
|
|
Calculated Results after
|
|
|
是否共享原始数据: IPD sharing |
Yes |
|
共享原始数据的方式(说明:请填入公开原始数据日期和方式,如采用网络平台,需填该网络平台名称和网址): |
发表论文 |
|
The way of sharing IPD”(include metadata and protocol, If use web-based public database, please provide the url): |
The data of the study will be conditionally shared when the study is published. |
|
数据采集和管理(说明:数据采集和管理由两部分组成,一为病例记录表(Case Record Form, CRF),二为电子采集和管理系统(Electronic Data Capture, EDC),如ResMan即为一种基于互联网的EDC: |
CRF |
|
Data collection and Management (A standard data collection and management system include a CRF and an electronic data capture: |
CRF |
|
数据与安全监察委员会: Data and Safety Monitoring Committee: |
暂未确定/Not yet |