Clinical Research Directory

Locally Optimised Contouring With AI Technology for Radiotherapy

LOCATOR is a multicentre phase II randomised clinical trial that is looking at the process of contouring in radiation treatment for breast cancer patients. This study looks at whether contouring aided by artificial intelligence (AI) is comparable in quality to that of contouring done completely manually by a radiation oncologist. We are also looking at whether AI assisted contouring saves radiation oncologists time when compared to fully manual contouring. LOCATOR uses the LOCATOR software which is an in-house software developed locally and trained on local data.

Evaluation of Left Ventricular Ejection Fraction Using an Accelerated Cardiac Cine-MRI Sequence With Deep Learning-based Image Reconstructions

Left ventricular hypertrophy (LVH) is a common condition that may result from hypertension, hypertrophic cardiomyopathy, aortic valve stenosis, or certain metabolic disorders. Cardiac imaging is essential for diagnosis, prognostic assessment, and quantification of cardiac function. While transthoracic echocardiography remains widely used, it is limited by acoustic window dependence and inter-observer variability. Cardiovascular Magnetic Resonance (CMR) imaging currently serves as the reference standard for measuring left ventricular ejection fraction (LVEF), cardiac volumes, and tissue characterization. However, conventional cine-CMR sequences require repeated breath-holds, which are often challenging for elderly or dyspneic patients, generating respiratory motion artifacts that compromise image quality. Accelerated cine-CMR sequences with deep learning-based image reconstructions offer a promising alternative by significantly reducing acquisition time while preserving image quality. This study aims to evaluate whether these accelerated cine-CMR sequences provide LVEF measurements concordant with conventional cine-CMR sequences, with potential to improve patient comfort and reduce examination time.

IDEAL Study: Blinded RCT for the Impact of AI Model for Cerebral Aneurysms Detection on Patients' Diagnosis and Outcomes

This study (IEDAL study) intends to prospectively enroll more than 6450 patients who will undergo head CT angiography (CTA) scanning in the outpatient clinic. It will be carried out in 21 hospitals in more than 10 provinces in China. The patient's head CTA images will be randomly assigned to the True-AI and Sham-AI group with a ratio of 1:1, and the patients and radiologists are unaware of the allocation. The primary outcomes are sensitivity and specificity of detecting intracranial aneurysms. The secondary outcomes focus on the prognosis and outcomes of the patients.

Deep Learning for Automated Discrimination Between Stage T1-T2 and T3 Renal Cell Carcinoma on Contrast-Enhanced CT

This study aims to develop and validate a contrast-enhanced CT-based deep-learning model for automatic and accurate preoperative discrimination between T1-T2 and T3 renal cell carcinoma. By quantifying the model's diagnostic performance on an independent test set-using AUC, sensitivity, specificity, positive/negative predictive values, and decision-curve analysis-we will establish a decision-support tool that can be seamlessly integrated into clinical PACS, thereby reducing staging errors, refining surgical planning, and improving patient outcomes.

Predictive Performance of a Generative Model for Corneal Tomography After ICL Implantation

To evaluate the efficacy of a corneal tomography Imaging model in predicting postoperative vault based on preoperative corneal topography in Implantable Collamer Lens (ICL) surgery.

Diagnostic Performance of an AI-based Model for TCM Constitution Classification Using Ophthalmic Imaging

To evaluate the diagnostic performance of a multimodal deep learning model for identifying biased Traditional Chinese Medicine (TCM) constitutions using ophthalmic imaging

Construction of a Deep Learning-Based Precise Diagnostic Framework for Bladder Tumors Using Ultrasound: A Multicenter, Ambispective Cohort Study

This study aims to develop an ultrasound image-based deep learning system to enable automatic segmentation, T-staging, and pathological grading prediction of bladder tumors. It seeks to enhance the objectivity, accuracy, and efficiency of bladder cancer diagnosis, reduce reliance on physician experience, and provide support for precision medicine and resource optimization.

Deep Learning Model Predicts Pathological Complete Response of Esophageal Squamous Cell Carcinoma Following Neoadjuvant Immunochemotherapy

This study aims to develop and validate a deep learning model to predict pathological complete response (pCR) in patients with esophageal squamous cell carcinoma who have undergone neoadjuvant immunochemotherapy. Clinical, imaging, and pathological data from previously treated patients will be collected and analyzed. The model is expected to assist in predicting treatment outcomes and guide personalized therapeutic strategies.

CT and Endoscopic Biopsy Image-Based Deep Learning for Predicting Left Recurrent Laryngeal Nerve Lymph Node Metastasis in Esophageal Cancer

The goal of this observational study is to develop a predictive model for left recurrent laryngeal nerve (RLN) lymph node metastasis using deep learning algorithms. The model will be developed using clinical data from previous esophageal cancer surgeries, including preoperative CT imaging, and histopathological images from gastroscopic biopsies. The model will also be validated through prospective clinical trials to guide the intraoperative lymph node dissection, thereby reducing postoperative risks of RLN injury.

Bladder Cancer Staging and Prediction of New Adjuvant Chemotherapy Efficacy Based on Deep Learning and Transfer Learning in Ultrasound-Magnetic Resonance-Pathology Multimodal Multiscale

Bladder cancer is the most common malignant tumor of the urinary system. The presence or absence of muscle invasion in early bladder cancer is an independent prognostic factor. The involvement of muscle invasion affects the choice of surgical methods and treatment. Preoperatively, the precise assessment of bladder cancer staging has important practical value. A more accurate preoperative assessment of bladder cancer staging can reduce overtreatment and provide a favorable basis for clinicians to choose more reasonable and effective surgical methods. Clinically, there has been a longstanding desire to diagnose the staging of bladder cancer through a simple, convenient, effective, and non-invasive examination. As relevant research progresses, a multi-omics diagnostic model will be beneficial in improving diagnostic efficiency. This project aims to establish a multi-omics artificial intelligence system based on deep learning and transfer learning to accurately diagnose the staging of bladder cancer and predict the efficacy of neoadjuvant chemotherapy. This system will assist in clinical treatment decision-making.

Deep Learning Using Chest X-Rays to Identify High Risk Patients for Lung Cancer Screening CT

The goal of this clinical trial is to evaluate whether an AI tool that alerts providers to patients at high 6-year risk of lung cancer based on their chest x-ray images will improve lung cancer screening CT participation. The main question it aims to answer is: Does the AI tool improve lung cancer screening CT participation at 6 months after the baseline outpatient visit The intervention is an alert to the provider to discuss lung cancer screening CT eligibility, for patients considered at high risk of lung cancer based on CXR-LC AI tool. If there is a comparison group: Researchers will compare intervention and non-intervention arms to determine if lung cancer screen CT participation increases.

Validation of the Prognostic Impact of a Retinal Photograph-based Cardiovascular Disease Risk Stratification System in de Novo HFrEF

"Despite significant advances in pharmacologic and device-based therapies, heart failure (HF) remains a major public health burden, with persistently high rates of hospitalization, impaired quality of life, and excess mortality-often exceeding those of leading malignancies. Prognosis in HF is shaped by its underlying etiology: ischemic HF often responds to revascularization strategies, whereas non-ischemic HF, particularly due to idiopathic or genetic cardiomyopathies, demonstrates highly variable outcomes and limited responsiveness to guideline-directed medical therapy (GDMT). Although left ventricular reverse remodeling (LVRR) is associated with favorable outcomes, only 40-50% of non-ischemic HF patients achieve meaningful LVRR with GDMT alone. In this context of therapeutic uncertainty and prognostic heterogeneity, there is a critical need for novel, non-invasive risk stratification tools. Retinal imaging offers a unique advantage, enabling direct, in vivo visualization of systemic microvascular and neurovascular integrity. Prior work from our group has demonstrated that deep learning algorithms applied to retinal fundus photographs can estimate physiologic and metabolic markers-including CAC scores-and predict future cardiovascular events. The Reti-CVD scoring system, derived from these models, has been externally validated in independent populations. In the present study, we aim to evaluate the prognostic utility of the Reti-CVD model in a cohort of patients with newly diagnosed HF and reduced ejection fraction. Specifically, we will assess whether retinal-derived risk scores at baseline are associated with adverse clinical outcomes, including cardiovascular events and all-cause mortality, and whether prognostic performance varies according to HF etiology."

AI-Assisted Smart Interactive Healthcare Robot

To address workforce shortages and increasing workloads in nursing, technological solutions and AI-powered robots for ward navigation have been introduced. However, limitations remain in their application to clinical care. This study aims to develop and test a programming framework for an AI-assisted nursing care robot ("E-Nursing Assistant") to reduce nurses' workload and improve the efficiency and quality of care.

The Construction and Effect Verification of a Deep Learning-based Automated Semantic Segmentation Model for Medical Imaging

Hepatocellular Carcinoma(HCC) is a common disease in China, ranking as the fourth most prevalent malignant tumor and the third leading cause of cancer-related deaths in the country. Along with other liver, biliary, pancreatic, and splenic diseases, it poses a serious threat to the lives and health of the Chinese population. Precise organ resection techniques, centered around accurate preoperative imaging and functional assessment as well as meticulous surgical operations, have become the mainstream in hepatobiliary surgery in the 21st century. These techniques require precise dissection of intrahepatic blood vessels, the biliary system, and the pancreatic-splenic duct system to achieve an optimal balance between eradicating lesions and preserving the normal function of the organs while minimizing trauma to the body. Precise tissue resection via laparoscopy is a prerequisite for successful hepatobiliary surgery. Addressing how to assist surgeons in performing surgeries more safely and effectively, as well as how to enhance learning outcomes during training, are pressing issues that need to be resolved. Efficient learning and analysis of surgical videos may help improve surgeons' intraoperative performance. In recent years, advancements in artificial intelligence (AI) have led to a surge in the application of computer vision (CV) in medical image analysis, including surgical videos. Laparoscopic surgery generates a large amount of surgical video data, providing a new opportunity for the enhancement of laparoscopic surgical CV technology. AI-based CV technology can utilize these surgical video data to develop real-time automated decision support tools and surgical training systems, offering new directions for addressing the shortcomings of laparoscopic surgery. However, the application of deep learning models in surgical procedures still has some shortcomings. Based on this, the present study aims to conduct a retrospective analysis of cases involving laparoscopic hepatobiliary and pancreatic surgeries performed at Zhujiang Hospital, Southern Medical University, between 2017 and 2024. The goal is to investigate the recognition and validation of deep learning models for classifying surgical phase images in medical imaging, as well as for semantic segmentation of anatomical structures, surgical instruments, and surgical gestures, including abdominal CT and MRI.

Artificial Intelligence in Detecting Cardiac Function

The Korotkoff Sounds（KS）, which have been in use for over a century, are widely regarded as the gold standard for measuring blood pressure. Furthermore, their potential extends beyond diagnosis and treatment of cardiovascular disease; however, research on the KS remains limited. Given the increasing incidence of heart failure (HF), there is a pressing need for a rapid and convenient prehospital screening method. In this study, we propose employing deep learning (DL) techniques to explore the feasibility of utilizing KS methodology in predicting functional changes in cardiac ejection fraction (LVEF) as an indicator of cardiac dysfunction.

Ga-68 Dolacga PET Scan in HCC Under RFA

This study aims to investigate the use of Ga-68 Dolacga PET scan technology to assess treatment response and liver function changes in patients of early-stage liver cancer receiving RFA. The main questions it aims to answer are: 1. How to assess treatment response and liver function changes in hepatocellular carcinoma patients undergo RFA via Ga-68 Dolacga PET scan? 2. Compared with computed tomography (CT) scans, how effective is Ga-68 Dolacga PET scan for treatment response assessment? 3. What is the correlation between Ga-68 Dolacga PET scan findings and patient treatment outcomes by tracking liver function and tumor recurrence after RFA? Participants will: 1. Undergo Ga-68 Dolacga PET scans and computed tomography before and one month after RFA treatment, followed by monitoring every three months thereafter. 2. Total liver functional volume and residual liver functional volume are obtained from Ga-68 Dolacga PET scan

Artificial Intelligence for Screening of Multiple Corneal Diseases

This study developed a deep learning algorithm based on anterior segment images and prospectively validated its ability to identify corneal diseases.The effectiveness and accuracy of this algorithm was evaluated by sensitivity, specificity, positive predictive value, negative predictive value, and area under curve.

Development and Demonstration of Intelligent Assessment Based on Multi-modal Information Fusion for Tumor Risk and Diagnosis and Treatment

To improve the accuracy of risk prediction, screening and treatment outcome of cancer, we aim to establish a medical database that includes standardized and structured clinical diagnosis and treatment information, image features, pathological features, and multi-omics information and to develop a multi-modal data fusion-based technology system using artificial intelligence technology based on database.

Implementation of Surgical Safety and Intraoperative Metastasis Identification Through Deep Learning: Multicentric Video Collection for Minimally Invasive Sentinel Lymph Node Dissection in Uterine Malignancies

The loco-regional metastatic or non-metastatic status of lymph nodes (LN) is critical for the therapeutic management of most cancer patients. Indeed, the presence or absence of lymphatic metastasis is essential for the accurate staging of the disease and strongly influence the prognosis and adjuvant treatment regimens. An important revolution in oncological surgery has been the introduction of the concept of sentinel lymph node (SLN) biopsy to reduce the complications of extensive loco-regional lymphadenectomies. SLN identification through ICG- based near-infrared fluorescence (NIR) cervical injection and its dissection is now recommended by European guidelines to stage uterine malignancies (endometrial and cervical cancers). However, SLN procedures have several limitations. In 11.2% of cases intra- or postoperative complications are reported due to anatomical structures injuries (vessels, nerves and lymphatic channels disruptions). Common mistakes, especially when the learning curve is not completed (at least 40 procedures), include mapping failure (25%) and removal of second/third-level nodes and/or empty nodes packets (8-14%). Additionally the intraoperative accuracy of frozen section is still far to be adequate with only the 65% of SLN metastasis detection. These limitations are a result of the lack of precision of current SLN localization and analysis as well as of the overall difficulty of visualizing lymph nodes and other critical structures in the retroperitoneum. Currently, studies on the safety of surgical procedures are based on perioperative clinical information and postoperative reports written by the surgeons themselves. Today, videos guiding minimally invasive surgical interventions allow for objective documentation of the procedure and provide opportunities to explore solutions for enhancing safety in the operating room. With an increasing use of endoscopic systems across different specialties, there is a need for standardization of training, assessment, testing and sign-off as a competent surgeon in order to improve patient safety. In laparoscopic lymph node dissection in endometrial and cervical cancer, a standardize stepwise approach to the procedure is highly recommended, by identifying key anatomic landmarks and structures, in various scenarios, that could prevent vascular, nervous and ureters injuries and enhance the mapping rate. Therefore, quantifying and studying intraoperative events such as the rate of achieving the right space dissections and anatomic structures visualization as a recommended step for safety and proficiency, would enable the examination of how best to implement guideline recommendations and seek new solutions to reduce operative risks. These videos could be utilized to train and validate artificial intelligence (AI) algorithms, with the potential to assist surgeons in the operating room and make the procedures safer. Additionally, the visual information (ICG intensity) could hide data that the AI can analyze and correlate with anatomopathological reports. By the integration of AI tool with laparoscopic/robotic platform it is possible to enhance MIS video streams in real time with surgical phases detection, events recognition, ICG signal intensity, anatomical structure identification and auto-targeting

Deep Learning for Preoperative Pulmonary Assessment in Thoracic CT

The trial was designed as a single-centre, non-interventional prospective observational study to utilize deep learning technology combined with computed tomography (CT) images to precisely predict the pulmonary function indicators of thoracic surgery preoperative patients.

A Prospective Study: Smart Phone Application for Measure Serum Bilirubin Through Sclera Images

The primary efficacy endpoints are the standard deviation and coefficient of determination (R2) between predicted and actual values for the bilirubin regression model, and the grading accuracy for the jaundice severity classification model. The secondary efficacy endpoint is the mean percentage error between predicted and actual bilirubin values. There are no relevant safety risks. Statistical differences for categorical variables (e.g., jaundice grading evaluation indicators) will be analyzed using the chi-square test or Fisher's exact probability test. For continuous variables (e.g., bilirubin prediction evaluation indicators), t-tests (normal distribution) or non-parametric tests (non-normal distribution) will be used. The 95% confidence interval for jaundice grading accuracy will be calculated using the Wilson method. The study duration is estimated to be 3 months.

Artificial Intelligence-enabled ECG Detection of Congenital Heart Disease in Children: a Novel Diagnostic Tool

Congenital heart disease (CHD) is the most common congenital disease in children. The early detection, diagnosis and treatment of CHD in children is of great significance to improve the prognosis and reduce the mortality of children, but the current screening methods have limitations. Electrocardiogram (ECG), as an economical and rapid means of heart disease detection, has a very important value in the auxiliary diagnosis of CHD.Big data and deep learning technologies in artificial intelligence (AI) have shown great potential in the medical field. The advent of the big data era provides rich data resources for the in-depth study of CHD ECG signals in children. The development of deep learning technology, especially the breakthrough in the field of image recognition, provides a strong technical support for the intelligent analysis of electrocardiogram. The particularity of children electrocardiogram requires the development of a special algorithm model. At present, the research on the application of deep learning models to identify children's electrocardiograms is limited, and the training and verification from large data sets are lacking. Based on the Chinese Congenital Heart Disease Collaborative Research Network, this project aims to integrate data and deep learning technology to develop a set of intelligent electrocardiogram assisted diagnosis system (CHD-ECG AI system) suitable for children with CHD, so as to improve the early detection rate of CHD and improve the efficiency of congenital heart disease screening.

Deep-learning For Ultrasound Classification of Anterior Talofibular Ligament Injury

Ultrasound (US) is a more cost-effective, accessible, and available imaging technique to assess anterior talofibular ligament (ATFL) injuries compared with magnetic resonance imaging (MRI). However, challenges in using this technique and increasing demand on qualified musculoskeletal (MSK) radiologists delay the diagnosis. Using datasets from multiple clinical centers, the investigators aimed to develop and validate a deep convolutional network (DCNN) model that automates classification of ATFL injuries using US images with the goal of providing interpretable assistance to radiologists and facilitating a more accurate diagnosis of ATFL injuries. The investigators collected US images of ATFL injuries which had arthroscopic surgery results as reference standard form 13 hospitals across China;Then the investigators divided the images into training dataset, internal validation dataset, and external validation dataset in a ratio of 8:1:1; the investigators chose an optimal DCNN model to test its diagnostic performance of the model, including the diagnostic accuracy, sensitivity, specificity, F1 score. At last, the investigators compared the diagnostic performance of the model with 12 radiologists at different levels of expertise.

Deep-learning Enabled Ultrasound Diagnosis of Anterior Talofibular Ligament Injury

Ultrasound (US) is a more cost-effective, accessible, and available imaging technique to assess anterior talofibular ligament (ATFL) injuries compared with magnetic resonance imaging (MRI). However, challenges in using this technique and increasing demand on qualified musculoskeletal (MSK) radiologists delay the diagnosis. The investigators have already developed a deep convolutional network (DCNN) model that automates detailed classification of ATFL injuries. The investigators hope to use the DCNN in real-world clinical setting to test its diagnostic accuracy.