This retrospective + prospective, non-interventional study aims to develop and evaluate artificial intelligence methods for the detection, pathological subtyping, and histological grading of renal tumors using magnetic resonance imaging (MRI). Approximately 900 adult patients with available preoperative renal MRI examinations and postoperative pathological results will be included. The pathological findings will be used as the reference standard for model development and evaluation. In addition to MRI data, selected demographic, clinical, and laboratory information may be incorporated to improve model performance. The study will not change participants' diagnosis, treatment, or follow-up, and no additional examinations or interventions will be required. All study data will be de-identified before analysis. The ultimate goal is to develop an MRI-based intelligent diagnostic approach that may assist clinicians in the preoperative assessment and individualized management of patients with renal tumors.
Patients aged 18 years or older.
Patients diagnosed with a renal tumor.
Availability of preoperative renal magnetic resonance imaging examinations.
Availability of a corresponding pathological diagnosis, including pathological subtype and, where applicable, histological grade.
Absence of renal magnetic resonance imaging data.
Absence of a corresponding pathological diagnosis or insufficient pathological subtype or grading information.
Magnetic resonance images that cannot be retrieved, opened, or read.
Poor image quality that precludes reliable image annotation or artificial intelligence analysis.
This retrospective + prospective, hospital-based study includes adult patients with renal tumors who underwent preoperative renal MRI as part of routine clinical care and had corresponding pathological diagnoses. Eligible cases include benign and malignant renal tumor subtypes. Existing de-identified MRI, pathological, demographic, clinical, and laboratory data are collected for artificial intelligence model development and evaluation. No healthy volunteers are included, and no additional examinations, treatments, or study-specific follow-up are performed.
Case-only
Other
MRI-Based Artificial Intelligence Analysis
Diagnostic testExisting preoperative multisequence renal MRI images, including T1-weighted imaging, T2-weighted imaging, diffusion-weighted imaging, apparent diffusion coefficient imaging, fat-suppressed imaging, and contrast-enhanced imaging when available, were retrospectively analyzed using artificial intelligence and deep learning methods. The models were developed to detect and segment renal tumors and to predict pathological subtype and histological grade. Postoperative pathological findings were used as the reference standard. No additional MRI examination or diagnostic procedure was performed for the study.
Accuracy of MRI-Based Artificial Intelligence for Pathological Subtyping of Renal Tumors
The pathological subtype predicted by the MRI-based artificial intelligence model will be compared with the postoperative pathological diagnosis as the reference standard in the held-out test dataset. Accuracy will be calculated as the number of correctly classified renal tumors divided by the total number of renal tumors evaluated. Classification performance for individual pathological subtypes will also be summarized using sensitivity, specificity, and F1 score, where applicable.
Accuracy of MRI-Based Artificial Intelligence for Histological Grading of Malignant Renal Tumors
The histological grade predicted by the MRI-based artificial intelligence model will be compared with the postoperative pathological grade as the reference standard. Histological grading will be assessed according to the four-tier World Health Organization/International Society of Urological Pathology grading system. Accuracy will be calculated as the number of malignant renal tumors with correctly predicted histological grade divided by the total number of malignant renal tumors evaluated.
Performance of the Artificial Intelligence Model for Renal Tumor Detection
The ability of the artificial intelligence model to detect and localize renal tumors on multisequence MRI will be evaluated by comparing model-generated tumor locations with expert manual annotations. Detection performance will be summarized using sensitivity and the proportion of correctly localized renal tumors.
Accuracy of Artificial Intelligence-Based Renal Tumor Segmentation
Artificial intelligence-generated renal tumor segmentations will be compared with pixel-level manual annotations prepared under the supervision of experienced physicians. Segmentation agreement will be evaluated using the Dice similarity coefficient.
Click on the lead sponsor to view all of their trials.
Cancer Institute and Hospital, Chinese Academy of Medical Sciences
Lead sponsor
Peking University People's Hospital
Collaborator
Shanxi Province Cancer Hospital
Collaborator
Chinese PLA General Hospital
Collaborator
RenJi Hospital
Collaborator
Cancer Hospital Chinese Academy of Medical Science, Shenzhen Center
Collaborator
