In August 2026, the Medical Ethics Subcommittee of the National Science and Technology Ethics Committee issued the Ethical Guidelines for Artificial Intelligence Research in Medical Imaging, providing ethical guidance for algorithm research, technology development, clinical validation, and translational application of artificial intelligence (AI) in medical imaging. Focusing on studies of diagnosis and decision support, this article briefly introduces the basic elements of ethical review, analyzes the effects of data trustworthiness and model trustworthiness on research, explores potential issues such as clinical translation and performance drift that may arise after technologies enter clinical application, and also addresses collaborative governance issues in whole-process management, including accountability and traceability and conflict-of-interest management. By examining how the Guidelines can be appropriately applied to whole-process research management, and how research conclusions and the scope of translation can be aligned with the available evidence, this article provides a reference for protecting the rights and interests of research participants and patients, and promoting responsible innovation.
Objective To compare the value of radiomic features derived from three diffusion-weighted imaging (DWI) models in predicting overall survival (OS) in patients with esophageal squamous cell carcinoma (ESCC). Methods Ninety patients with pathologically confirmed ESCC were retrospectively included (mean age, 64.9±7.2 years; median follow-up, 42 months). DWI was acquired at 3.0 T using 12 b values (0-2 000 s/mm²). Tumor regions of interest (ROIs) were delineated on T2-weighted images (T2WI), registered to DWI-derived parametric maps from the monoexponential apparent diffusion coefficient (Mono-ADC), fractional-order calculus (FROC), and continuous-time random-walk (CTRW) models, and used for radiomic feature extraction. Univariate Cox regression and least absolute shrinkage and selection operator (LASSO)-Cox regression were used to construct a radiomics score (Rad-score) for each of the three models. Patients were divided into high- and low-risk groups according to the median Rad-score (45 patients in each group). Predictive performance was evaluated using the concordance index (C-index) and time-dependent receiver operating characteristic (ROC) curves; the area under the curve (AUC) was calculated, and differences in AUC among the Rad-scores were compared using the DeLong test. Survival analysis was performed using the Kaplan-Meier method, and between-group differences were assessed using the log-rank test. Univariate and multivariable Cox regression analyses were used to identify independent prognostic factors. Results LASSO-Cox regression selected 2, 3, and 3 features to construct the Rad-scores for the Mono-ADC, FROC, and CTRW models, respectively. The C-index of the Rad-score_CTRW model was 0.785, higher than those of the Rad-score_FROC model (0.735) and the Rad-score_ADC model (0.681). For predicting 3-year survival, the AUC of the Rad-score_CTRW model was 0.761, higher than those of the Rad-score_FROC model (0.703; Z=2.07, P=0.038) and the Rad-score_ADC model (0.652; Z=3.65, P<0.001). Only risk stratification based on the Rad-score_CTRW model significantly differentiated OS between the high- and low-risk groups (P=0.001). Multivariable Cox regression analysis showed that all three Rad-scores were independent predictors of OS. Conclusion Radiomic features derived from the CTRW model showed better performance in predicting OS in patients with ESCC than those derived from the FROC and Mono-ADC models.
Objective To analyze the reproducibility of radiomic features extracted from pulmonary ground-glass nodules (GGNs) on low-dose and conventional-dose CT. Methods A total of 292 patients who underwent low-dose CT screening for pulmonary nodules and additional conventional-dose CT scanning of the GGN region were retrospectively included. In total, 306 GGNs with a maximum diameter of 5-30 mm were included. Two physicians manually segmented GGNs on both types of CT images, and 107 radiomic features (shape features, first-order features, and texture features) were extracted. The intraclass correlation coefficient (ICC) was used to evaluate interobserver agreement in radiomic feature extraction. The radiation doses [volume CT dose index (CTDIvol), dose-length product (DLP), and effective dose (ED)] were compared between the two scanning protocols using the paired t-test. The concordance correlation coefficient (CCC) was used to evaluate the reproducibility of radiomic features extracted by the two scanning protocols. Results The interobserver and intraobserver agreement for radiomic feature extraction was high (all ICC>0.96). The CTDIvol, DLP, and ED were lower with low-dose CT than with conventional-dose CT (all P<0.05). Of the 107 radiomic features, 48 (44.9%) showed good reproducibility (CCC≥0.85), whereas 59 (55.1%) showed poor reproducibility (CCC<0.85). First-order features had the highest CCC (0.90±0.08), with 77.8% (14/18) showing good reproducibility, while shape and texture features had lower CCCs (0.79±0.24 and 0.79±0.14, respectively), and 65.3% (49/75) of texture features showed poor reproducibility. Conclusion Among the GGN radiomic features extracted from low-dose and conventional-dose CT, first-order features showed good reproducibility, whereas texture features showed poor reproducibility; features with good reproducibility should be preferentially selected for low-dose CT radiomics analysis.
Objective To investigate the value of spectral CT parameters combined with clinicopathological characteristics for predicting postoperative disease-free survival (DFS) risk in patients with locally advanced rectal cancer (LARC). Methods This retrospective study included 198 patients with LARC who underwent preoperative spectral CT examination followed by radical surgery. Patients were randomly divided into the training cohort (n=138) and validation cohort (n=60) at a ratio of 7∶3. Venous-phase normalized iodine concentration (NICVP) and delayed-phase normalized iodine concentration (NICDP) were measured using the maximum cross-sectional region-of-interest (ROI) method and the whole-tumor ROI method. Interobserver agreement and the predictive performance of imaging models derived from the two ROI methods were compared. Cox regression analyses were subsequently performed using parameters obtained with the selected method, and clinical, imaging, and imaging-clinical combined models were developed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and DeLong test. The best-performing model was further used for risk stratification and nomogram construction. Results The interobserver ICCs for NICVP and NICDP measured using the whole-tumor ROI method were 0.95 and 0.82, respectively, higher than those obtained using the maximum cross-sectional ROI method (0.79 and 0.69, respectively). The corresponding imaging model also achieved a higher AUC (0.83 vs. 0.75, P<0.05). Univariate Cox regression showed that histologic differentiation, T stage, N stage, lymphovascular invasion, perineural invasion, NICVP, and NICDP were significantly associated with DFS. Multivariable Cox regression analysis further identified histological low differentiation, lymphovascular invasion, perineural invasion, and NICVP as independent prognostic factors (P<0.05). The imaging-clinical combined model demonstrated superior predictive performance compared with the clinical model in both the training and validation cohorts (AUCs: 0.92 vs. 0.87, 0.84 vs. 0.75, respectively). After risk stratification based on the combined-model risk score, DFS differed significantly between the high- and low-risk groups (P<0.001). A nomogram based on this model enabled individualized prediction of 2-year DFS, with a C-index of 0.844. Conclusions The whole-tumor ROI method demonstrated higher interobserver agreement. Spectral CT quantitative parameters combined with clinicopathological characteristics can improve predictive performance in patients with LARC and may provide a reference for postoperative risk stratification.
Objective To investigate CT imaging differences between abdominopelvic Castleman disease (CD) and paraganglioma (PGL), identify imaging features with differential value, construct a diagnostic model, and evaluate its value for preoperative differential diagnosis. Methods A total of 57 patients with abdominopelvic CD or PGL, including 24 patients with CD and 33 patients with PGL, whose diagnoses were confirmed by postoperative histopathology, were retrospectively included. All patients underwent unenhanced and three-phase contrast-enhanced CT examinations. Multiphase CT attenuation values and morphological features of the lesions were assessed. The Mann-Whitney U test and the chi-square test were used to compare the two groups. Variables showing statistically significant differences in the univariate analysis were entered into multivariable logistic regression. Forward stepwise selection was used to select variables and construct the differential diagnostic model. Receiver operating characteristic (ROC) curve analysis was performed to evaluate model performance. Results Venous-phase CT attenuation and the difference between venous-phase and unenhanced CT attenuation values (ΔHU) were higher in the PGL group than in the CD group.Perilesional fascial reaction, satellite lesions, and the rim sign were more frequent in the CD group than in the PGL group (all P<0.05). A combined model incorporating ΔHU, perilesional fascial reaction, satellite lesions, and the rim sign yielded an area under the ROC curve (AUC) of 0.914 (95%CI: 0.809-0.972), which was higher than that of venous-phase CT attenuation alone (AUC=0.693) or ΔHU alone (AUC=0.711). At a predicted probability cutoff of P(CD)=0.50, the combined model achieved the highest sensitivity (0.792), specificity (0.939), and accuracy (0.877). Conclusion The logistic regression model based on ΔHU, perilesional fascial reaction, satellite lesions, and the rim sign had good performance in differentiating abdominopelvic CD from PGL. This model may provide imaging-based information for preoperative differential diagnosis and perioperative management.
Objective To investigate the predictive value of computed tomography (CT) features at the transition zone in adhesive small bowel obstruction (ASBO) for the outcome of conservative treatment, and to provide imaging-based evidence for clinical decision-making. Methods Clinical and CT imaging data of 235 patients with ASBO were retrospectively analyzed. Based on the treatment modality, patients were divided into the conservative treatment success group (n=170) and the non-success group (n=65). Two radiologists independently analyzed the location of the transition zone, adhesion morphology at the transition zone, degree of obstruction, and small bowel feces sign on CT images. Interobserver agreement was assessed using Cohen’s Kappa coefficient. Comparisons of general characteristics and CT features between the two groups were performed using the t test and chi-square test. Univariate and multivariate logistic regression analyses were used to identify CT features influencing the outcome of conservative treatment, and predictive models based on individual features and their combination were constructed. Receiver operating characteristic (ROC) curves were used to evaluate model performance, and the area under the curve (AUC) was calculated. Results Interobserver agreement for all evaluated CT features at the transition zone was excellent (all κ>0.85). The conservative treatment success group showed significantly higher proportions of small bowel feces sign and angled adhesion compared with the non-success group, whereas complete obstruction and band-like impression were more frequent in the non-success group (all P<0.05). Multivariate analysis revealed that small bowel feces sign (OR=3.307) was an independent protective factor for successful conservative treatment, while band-like impression (OR=0.192) and complete obstruction (OR=0.225) were independent risk factors (all P<0.05). Among the individual-feature models and the combined model incorporating all three features, the combined model demonstrated the highest predictive performance (AUC=0.764). Conclusion Identification of CT features at the transition zone in adhesive small bowel obstruction can effectively predict the outcome of conservative treatment, thereby facilitating optimization of clinical management strategies.
Head and neck CT angiography (CTA) is currently the first-line non-invasive imaging modality for diagnosing cerebrovascular diseases. Compared with conventional CT, photon-counting CT (PCCT) provides higher spatial and spectral resolution, enhanced multi-material decomposition capability, and a lower radiation dose, showing promising potential for head and neck vascular imaging. This review briefly describes the imaging principles and core technical advantages of PCCT, summarizes current strategies for optimizing acquisition and reconstruction parameters for PCCT-based head and neck CTA, and reviews advances in its clinical applications for visualizing small head and neck vessels, identifying high-risk plaques in the head and neck arteries, assessing intracranial aneurysms before and after treatment, and performing follow-up after head and neck vascular stent placement.
Epicardial adipose tissue (EAT) is involved in the pathophysiology of heart failure with preserved ejection fraction (HFpEF) through mechanical compression and the secretion of bioactive substances. Currently, commonly used imaging modalities for evaluating EAT include echocardiography, computed tomography (CT), and cardiac magnetic resonance (CMR). Quantitative assessment of EAT can be achieved by measuring parameters such as thickness, volume, density, and texture entropy. With advances in artificial intelligence (AI), deep learning-based automated segmentation and feature mining provide new approaches for quantitative EAT analysis. This review summarizes the pathophysiological mechanisms by which EAT is involved in HFpEF, the current status of imaging techniques for quantitative EAT assessment, and research advances in the application of AI to EAT analysis.
Lung adenocarcinoma (LUAD) exhibits marked biological and spatial heterogeneity. Accurate preoperative assessment of pathological invasiveness in early-stage LUAD, including high-grade components, lymph node metastasis, and microscopic invasive features, is critical for individualized surgical and adjuvant treatment planning. Habitat imaging decomposes tumors into spatial subregions with distinct biological characteristics, thereby enabling quantitative assessment of intratumoral heterogeneity. This technique has shown advantages in predicting high-grade histological subtypes, assessing lymph node metastasis status, and identifying microscopic invasive behaviors such as lymphovascular invasion, visceral pleural invasion, and spread through air spaces in early-stage LUAD. This review summarizes recent advances in habitat imaging for assessing invasiveness in early-stage LUAD.
Early diagnosis, precise pathological classification, and treatment response assessment of lung cancer are crucial for improving patient prognosis. Dual-layer spectral computed tomography (DLCT), a new-generation spectral CT technique, can simultaneously acquire high- and low-energy data in a single scan and generate multiple quantitative parameters, including iodine concentration, effective atomic number, virtual monoenergetic images, electron density, and slope of the spectral Hounsfield unit curve. These parameters provide a new quantitative imaging approach for the noninvasive assessment of lung cancer. This review introduces the basic imaging principles and technical advantages of DLCT，and focuses on recent advances in its multiparametric applications to early screening for lung cancer, differentiation of benign and malignant pulmonary nodules, pathological classification, lymph node metastasis assessment, treatment response monitoring, radiotherapy planning optimization, and prognosis prediction.Furthermore, the current limitations and future directions are also discussed.
Spectral CT can provide multiple types of spectral CT images and quantitative parameters. It can be used to differentiate lung cancer from benign pulmonary lesions and provide quantitative imaging evidence for histological subtyping, assessment of the degree of invasion and pathological grading, evaluation of local invasive features, and assessment of immune biomarker expression and driver gene mutations in lung cancer. The integration of spectral CT with radiomics, machine learning, and deep learning further expands its value in the precision diagnosis of lung cancer. This review summarizes research advances in spectral CT for the precision diagnosis of lung cancer and analyzes current issues and future directions.
Rectal cancer exhibits marked spatial tumor heterogeneity, which is an important contributor to treatment resistance and poor prognosis. Accurate assessment of tumor heterogeneity is essential for individualized diagnosis and treatment. Habitat imaging can noninvasively and quantitatively distinguish intratumoral subregions with distinct microenvironmental characteristics, providing a novel approach for assessing tumor heterogeneity. This review systematically summarizes the analytical workflow of MRI habitat imaging and its clinical utility in preoperative T staging of rectal cancer, prediction of aggressive pathological features, molecular classification of microsatellite instability, assessment of response to neoadjuvant chemoradiotherapy and immunotherapy, and risk stratification for metachronous distant metastasis. Future research directions are also discussed, including the establishment of multicenter standards, iterative optimization of intelligent algorithms, and integration of imaging with spatial multi-omics, to provide an imaging-based theoretical foundation for precision and individualized diagnosis and treatment of rectal cancer.
Osteosarcopenia (OS) is a common geriatric syndrome that can increase the risks of falls, fragility fractures, functional limitations, and mortality. Imaging methods, including dual-energy X-ray absorptiometry (DXA), CT, MRI, and ultrasonography, can assess bone mineral density, bone microarchitecture, muscle mass, and muscle quality in OS and can be used for diagnosis, risk identification, opportunistic screening, prognostic risk prediction, and imaging follow-up management. This review summarizes advances in the application of various imaging techniques to the identification and assessment of OS.
Parkinson’s disease (PD) is a neurodegenerative disorder primarily characterized by the degeneration of dopaminergic neurons and abnormal aggregation of α-synuclein (α-syn). Nuclear medicine molecular imaging enables the in vivo visualization and quantitative assessment of relevant molecular targets using specific imaging agents, providing accurate diagnosis and therapeutic evaluation in patients with PD. This review systematically summarizes research advances in imaging agents targeting the dopaminergic system [including dopamine transporter(DAT), vesicular monoamine transporter type 2 (VMAT2), dopamine receptor(DR), and dopamine analogs], α-syn, and non-dopaminergic targets, including neuroinflammation, mitochondrial dysfunction, cerebral glucose metabolism, cholinergic and serotonergic systems. Particular emphasis is placed on major advances in the development of novel α-syn probes, along with an analysis of their pharmacokinetic characteristics and prospects for clinical application.
Objective To analyze the CT and MRI features of early- and middle-stage atlantoaxial rheumatoid arthritis (RA) and atlantoaxial osteoarthritis (OA), and to investigate their value in differential diagnosis. Methods Sixty-two patients with early- and middle-stage RA diagnosed clinically and by imaging were retrospectively enrolled, including 31 patients with atlantoaxial involvement and 31 without atlantoaxial involvement. Patients with atlantoaxial involvement were further divided into two subgroups according to disease duration: <20 years (13 cases) and ≥20 years (18 cases). During the same period, 47 patients with atlantoaxial OA were enrolled. Clinical and imaging data from patients with atlantoaxial RA and OA were collected, and their CT and MRI features were analyzed. The Mann-Whitney U test, chi-square test, and Fisher exact test were used to compare clinical data between the RA groups with and without atlantoaxial involvement, clinical and imaging features between the atlantoaxial involvement group and the OA group, and the proportions of imaging features in atlantoaxial RA with different disease durations. Results The main imaging findings of early- and middle-stage atlantoaxial RA were anterior dislocation and vertical subluxation, atlantoaxial bone destruction and odontoid erosion, and synovial thickening. Atlantoaxial OA was mainly characterized by cartilage wear, osteophyte formation, and sclerosis. Disease duration and the proportion of patients with “high disease activity” on the 28-joint Disease Activity Score (DAS28) were higher in the RA group with atlantoaxial involvement than in the RA group without atlantoaxial involvement (all P<0.05). Patients in the RA group with atlantoaxial involvement were younger and had a lower proportion of atlantodental arthritis than those in the OA group, whereas the proportions of female patients, lateral atlantoaxial arthritis, anterior dislocation, vertical dislocation, erosion/irregularity of the posterior margin of the odontoid process, synovial thickening, and spinal canal narrowing were higher (all P<0.05). There were no statistically significant differences among disease-duration groups in the degree of odontoid destruction, the number of types of atlantoaxial subluxation, or the proportion of synovial thickening (all P>0.05). Conclusion The imaging features of early- and middle-stage atlantoaxial RA and OA differ to some extent. Erosion of the posterior margin of the odontoid process in RA has value for differential diagnosis and may provide a reference for clinical and imaging diagnosis.
