Objective To investigate the feasibility of 60 kVp ultra-low tube voltage combined with artificial intelligence iterative reconstruction (AIIR) in head and neck CT angiography (CTA). Methods A total of 120 patients undergoing head and neck CTA were prospectively enrolled and randomly divided into 4 groups (30 cases per group): Group A [100 kVp, 40 mL contrast agent, hybrid iterative reconstruction (HIR)], Group B (80 kVp, 40 mL contrast agent, HIR), Group C (60 kVp, 40 mL contrast agent, AIIR level 3 and 4, recorded as C3 and C4), and Group D (60 kVp, 20 mL contrast agent, AIIR level 3 and 4, recorded as D3 and D4). Subjective image quality scores (5-point scale), objective parameters [noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR)], radiation dose [CT dose index (CTDIvol), dose-length product (DLP), and effective dose (ED)] were compared among groups. Diagnostic accuracy was assessed using digital subtraction angiography (DSA) as the reference standard. Continuous variables were analyzed using one-way ANOVA or the Kruskal-Wallis test, categorical variables using the chi-square test, and interobserver agreement was assessed using Cohen’s Kappa test. Results Baseline characteristics were well balanced among the 4 groups (all P>0.05). The subjective image quality scores were ≥3 in all groups, with no significant differences among groups (all P>0.05), and inter-reader agreement was good (all κ≥0.675). Compared with Group A, image noise was significantly reduced in Group C3, C4, D3, and D4 (all P<0.05). while SNR and CNR were significantly increased (all P<0.05). No significant difference was found between Groups B and A (P>0.05). The mean CT values in the low-dose groups were comparable to or significantly higher than those in Group A. Thirteen patients underwent DSA examination, and comparison with DSA demonstrated a diagnostic accuracy of 100% for CTA in all groups. Compared with Group A, radiation dose was reduced by 44.7%, 77.3%, and 77.3% in Groups B, C, and D, respectively. Conclusion The 60 kVp head and neck CTA protocol combined with AIIR can maintain diagnostic-level image quality while reducing radiation dose by 77% and contrast agent volume by 50%.

Objective To investigate the application and value of ultra-low tube voltage (60 kVp) combined with an artificial intelligence iterative reconstruction (AIIR, United Imaging Healthcare uExceed R001) in CT angiography (CTA) of upper limb arteriovenous fistula. Methods Ninety patients with chronic renal failure undergoing upper limb arteriovenous fistula CTA were prospectively enrolled. Using a random number table, they wer randomly assigned to a conventional dose group (100 kVp, 1.0 mL/kg contrast medium, hybrid iterative reconstruction; n=45) and an ultra-low dose group (60 kVp, 0.5 mL/kg contrast medium, AIIR reconstruction; n=45). Two senior radiologists independently assessed subjective image quality and diagnostic confidence for anastomotic stenosis using a 4-point scale. Cohen’s Kappa was used to evaluate inter-observer agreement. The radiologists quantitatively analyzed the mean CT attenuation, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) at the subclavian, brachial artery, and radial artery levels. Radiation dose and contrast medium usage were compared between the two groups. Differences between the two groups were compared using the independent sample t-test, Mann-Whitney U test or chi-square test. Results There were no statistically significant differences in age, sex, body weight, and body mass index between the two groups (P>0.05). Image quality and fistula diagnostic confidence scores were ≥3 points, meeting clinical diagnostic requirements, with good interobserver agreement (Kappa>0.72). Quantitative measurements showed that the ultra-low dose group had significantly higher SNR and CNR values at all measured vascular levels compared with the conventional dose group (all P<0.001), including the subclavian artery, brachial artery, and radial artery. The effective radiation dose in the ultra-low dose group was significantly lower than that in the conventional dose group (1.28 mSv vs. 6.50 mSv, P<0.001), reduced by approximately 80%. Contrast medium volume decreased from 61.0 mL to 28.5 mL (53% reduction), and iodine load decreased from 21.35 g to 11.40 g (47% reduction). Conclusion Ultra-low dose combined with AIIR in upper limb arteriovenous fistula CTA can significantly reduce radiation dose and contrast medium usage while maintaining or improving image quality and diagnostic confidence. It demonstrates potential clinical value, particularly suitable for high-frequency follow-up in dialysis patients.

Objective To investigate the feasibility of 60 kVp CT angiography (CTA) combined with a cardiac deep learning reconstruction algorithm (CardioBoost) for children with congenital heart disease (CHD). Methods A total of 160 children with clinically confirmed CHD who were scheduled for preoperative cardiac CTA were prospectively enrolled, with a median age of 8 months. The children were randomly assigned to a conventional-dose group and a low-dose group (80 patients in each group). In the conventional-dose group, hybrid iterative reconstruction (HIR) was used to generate Group A images (80 kVp, 400 mA, HIR). In the low-dose group, HIR and CardioBoost reconstruction were used to produce Group B1 images (60 kVp, 240 mA, HIR) and Group B2 images (60 kVp, 240 mA, CardioBoost reconstruction), respectively. Radiation doses between the standard-dose and the low-dose groups were compared, including volume CT dose index (CTDIvol), dose-length product (DLP), effective dose (ED), and size-specific dose estimate (SSDE). The subjective image quality scores (5-point scale) and objective image quality parameters of each cardiac chamber and great vessel, including image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR), were compared among Groups A, B1, and B2. Surgical findings were used as the gold standard to evaluate diagnostic accuracy. Quantitative data were compared using the Mann-Whitney U test and Kruskal-Wallis test, and categorical variables were compared using the chi-square test. Interobserver agreement between the two readers was assessed using Cohen’s Kappa test. Results The ED in the low-dose group was reduced by 81.3% compared with that in the conventional-dose group (P<0.05). The interobserver agreement for the subjective image quality scoring was good across all three groups (all κ≥0.78). The subjective image quality scores were all ≥3 in Groups A, B1, and B2. The subjective image quality score in Group B2 was significantly higher than that in Group B1 (P<0.05), but did not differ significantly from that in Group A (P>0.05). Compared with Group B1, Group B2 showed reduced image noise and increased SNR and CNR (all P<0.05), and the objective image quality indices were comparable to or better than those of Group A. The diagnostic accuracy for intracardiac malformations in Group B2 was significantly higher than that in Group B1 (P<0.05), but did not differ significantly from that in Group A (P>0.05). Conclusion For CTA in children with CHD, 60 kVp combined with the CardioBoost algorithm can significantly reduce radiation dose while maintaining image quality comparable to that of the conventional-dose protocol.

Objective To investigate the feasibility of using pharmacokinetic parameters derived from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to differentiate microvascular heterogeneity between tumor tissue and proximal tumor-distant tissue (PTD) in esophageal squamous cell carcinoma (ESCC). Methods A total of 154 patients with pathologically confirmed ESCC from two medical centers were prospectively enrolled. There were 128 patients from Center A were randomly assigned to a training set (102 cases) and an internal validation set (26 cases) at a ratio of 8∶2. Patients from Center B (26 cases) served as an external validation set. The training set was used for selection of pharmacokinetic parameters reflecting microvascular heterogeneity and for Logistic regression model construction, while the validation sets were used to assess the diagnostic performance of the model. Regions of interest (ROIs) were delineated in tumor tissue and PTD using medical imaging processing software developed by United Imaging Healthcare. Sensitivity analyses were performed to assess the robustness of the model by comparing ROI delineation results with and without exclusion of necrotic and cystic areas. The mean, standard deviation (SD), and coefficient of variation (CV) of the reflux rate constant (k_ep), volume transfer constant (K^trans), and extracellular extravascular volume fraction (v_e) were extracted. The Wilcoxon signed-rank test was used to compare differences in pharmacokinetic parameters between tumor tissue and PTD. Parameters with statistically significant differences were included in multivariate Logistic regression analysis. Receiver operating characteristic (ROC) curves were used to evaluate the discriminative performance of single parameters and multivariable models, and the area under the curve (AUC) was calculated. DeLong test was applied to compare differences in AUCs. Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated to assess reclassification ability of the model. Results In the training set, the mean of k_ep, SD of k_ep, mean of K^trans, mean of v_e, and CV of v_e showed significant differences between tumor tissue and PTD (all P<0.05). Multivariable Logistic regression analysis identified the mean of k_ep and the CV of v_e as independent predictors for differentiating tumor tissue from PTD (both P<0.05), and a multivariable model was constructed based on these two parameters. The AUC values of the multivariable model in the training, internal validation, and external validation sets were 0.835, 0.846, and 0.818, respectively. The DeLong test showed that the AUCs of the multivariable model were significantly higher than those of the mean of k_ep (all P<0.05), while there was no significant difference between the AUCs of the multivariable model and those of the CV of v_e (all P>0.05). The multivariable model performed better than the mean of k_ep and comparably to the CV of v_e, whereas NRI and IDI indicated that the multivariable model had superior reclassification ability (all P<0.05). Sensitivity analysis showed that whether necrotic or cystic regions were excluded from the ROIs or not, the discriminative performance of the model and the direction of parameter effects remained consistent. Conclusion The multivariable model based on the CV of v_e and the mean of k_ep demonstrates good performance in distinguishing ESCC tumor tissue from PTD.

Objective To investigate the value of an intratumoral-peritumoral habitat radiomics model based on contrast-enhanced CT for the preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Methods A total of 158 patients with pathologically confirmed HCC and known MVI status were retrospectively included. Clinical data and CT imaging features from arterial-phase and portal venous-phase contrast-enhanced upper abdominal scans were analyzed. The patients were randomly divided into a training set (n=110) and a testing set (n=48) at a ratio of 7∶3. Radiomics features were extracted from intratumoral, intratumoral-peritumoral, and intratumoral-peritumoral habitat subregions. Optimal features were selected using the least absolute shrinkage and selection operator (LASSO) algorithm. Univariate and multivariate logistic regression analysis were performed on clinical data to identify independent risk factors for predicting MVI. Models were constructed using the support vector machine (SVM) machine learning algorithm. Four models were developed using the training set: an intratumoral radiomics model, an intratumoral-peritumoral radiomics model, an intratumoral-peritumoral habitat radiomics model, and a clinical model. Model performance was evaluated in the testing set using the receiver operating characteristic (ROC) curve, the area under the curve (AUC), and net reclassification improvement (NRI). Calibration curves and decision curve analysis (DCA) were used to assess model calibration and clinical net benefit. Results Among the four constructed models, the intratumoral-peritumoral habitat radiomics model demonstrated the best predictive performance, with AUCs of 0.893 in the training set and 0.831 in the testing set. Compared with the intratumoral radiomics model, intratumoral-peritumoral radiomics model, and clinical model, the habitat radiomics model showed positive improvement in predictive ability (NRI>0). Calibration curves indicated good agreement between the predicted probabilities and the actual outcomes for the habitat radiomics model. Furthermore, compared with the clinical model, intratumoral radiomics model, and intratumoral-peritumoral radiomics model, the habitat radiomics model provided greater clinical net benefit. Conclusion The intratumoral-peritumoral habitat radiomics model based on contrast-enhanced CT exhibits excellent performance in predicting MVI in HCC patients. It has potential for noninvasive preoperative prediction of MVI and may provide a basis for personalized clinical treatment planning.

Objective To investigate the association between different skull fracture sites and the occurrence and complexity of intracranial hemorrhage in patients with traumatic brain injury (TBI), and to develop CT-based risk prediction models for intracranial hemorrhage using initial head computed tomography (CT) findings. Methods In this multicenter retrospective study, 4 700 TBI patients from five tertiary hospitals were enrolled. Based on the initial CT findings, patients were categorized into three groups: no intracranial hemorrhage (n=1 602), single intracranial hemorrhage (n=1 011), and multiple intracranial hemorrhages (n=2 087). Data on age, sex, scalp hematoma, midline shift, cerebral herniation, skull fracture sites, and intracranial hemorrhage types were recorded. Continuous variables were compared among groups using one-way analysis of variance (ANOVA), while categorical variables were compared using the chi-square test. Univariable analysis was used to assess the associations between skull fracture sites and various types of intracranial hemorrhage. Multivariable logistic regression was then performed to identify independent risk factors related to skull fracture sites for predicting the occurrence and complexity of intracranial hemorrhage.Prediction models for the presence/absence of intracranial hemorrhage and hemorrhage complexity were constructed accordingly, and model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), calibration curves, and decision curve analysis (DCA). Results Significant differences were observed among the three groups in age, sex, scalp hematoma, midline shift, cerebral herniation, and the distribution of skull fracture sites (all P<0.05). Cranial fractures were identified as risk factors for intracranial hemorrhage, with sphenoid fractures (OR=8.35) and temporal fractures (OR=6.93) showing the strongest associations. In the model predicting the presence or absence of intracranial hemorrhage, after incorporating clinical information and specific skull fracture sites, the model achieved an AUC of 0.895 (95%CI: 0.89-0.90), demonstrated good calibration, and DCA indicated a high net benefit within a threshold probability range of approximately 0.10-0.60. The prediction model for multiple intracranial hemorrhages had an AUC of 0.689 (95%CI: 0.67-0.71), showed acceptable calibration, and retained some clinical utility within certain threshold probability ranges. Furthermore, epidural hematoma (EDH) was significantly associated with temporal and sphenoid fractures, as well as high-risk combined cranial fracture patterns (all P<0.05). Conclusion Skull fracture sites are closely associated with the occurrence and complexity of intracranial hemorrhage in TBI patients. The risk prediction models for intracranial hemorrhage, constructed based on skull fracture sites and relevant clinical information, exhibit good discriminative ability and have certain clinical decision-making value, which may serve as a reference for early risk stratification of TBI patients in the emergency department.

Objective To evaluate the predictive value of multiple quantitative parameters derived from dual-energy computed tomography (DECT) for risk stratification of gastrointestinal stromal tumors (GISTs). Methods This retrospective study included 36 patients with surgically and pathologically confirmed GISTs, with a mean age of 62.2±10.8 years. Patients were categorized into a high-risk group (n=15) and a low-risk group (n=21) based on the number of mitotic nuclei. Clinical characteristics were assessed, and multi-quantitative parameters from DECT were measured, including iodine concentration (IC), normalized iodine concentration (NIC), fat fraction (FF)，electron density (Rho), effective atomic number (Zeff), and dual-energy index (DEI) on venous-phase fusion images. The Chi-square test or t-test was employed to compare differences in parameters between the two groups. Univariate and multivariate Logistic regression analyses were performed to identify independent predictors for high risk GIST, and a combined logistic regression model integrating multiple factors was constructed. Receiver operating characteristic (ROC) curves were used to assess the predictive performance of the model, and the area under the curve (AUC) was calculated. The DeLong’s test was used to compare difference in AUC values among models. Calibration curve and decision curve analysis (DCA) were used to evaluate model calibration and clinical applicability. Results Univariate Logistic regression identified tumor location, heterogeneous enhancement, necrosis/cystic degeneration, maximum diameter, venous-phase IC, and NIC as predictors of high-risk GIST (all P<0.05). Multivariate analysis revealed that maximal tumor diameter [odds ratio (OR)=1.59, P=0.012] and NIC (OR=1.08, P=0.014) were independent risk factors for high-risk GIST. The combined model constructed based on these two independent predictors achieved an AUC of 0.95, with a sensitivity of 93.8% and a specificity of 85.0%, showing significantly better predictive performance than each individual factor (all P<0.05). The calibration curve demonstrated good agreement between predicted probabilities and observed outcomes. DCA showed that the combined model yielded a higher net benefit within a threshold probability range of 0.05-0.73. Conclusion Multiple quantitative parameters derived from DECT can non-invasively predict the stratification of GISTs preoperatively, providing a reference for clinical treatment decision-making.

Glioblastoma (GBM) and solitary brain metastasis (SBM) exhibit similar conventional imaging features; however, their clinical treatment strategies differ significantly. Accurate differentiation between the two is therefore crucial for subsequent diagnosis and treatment. Deep learning, a branch of machine learning, can optimize multiple key steps in the image analysis workflow, including improving the efficiency of region-of-interest segmentation, accurately extracting imaging features, and constructing efficient fusion models, thus providing new solutions for differentiating GBM from SBM. Compared with traditional radiomic and machine learning, deep learning represents a more powerful and effective approach. This review systematically summarizes the current applications, technical progress, and challenges of deep learning in the differential diagnosis between GBM and SBM.

Functional recovery after ischemic stroke exhibits considerable interindividual variation that conventional clinical measures often fail to fully explain. Advances in artificial intelligence have enabled novel neuroimaging-based approaches. Metrics such as brain structure damage, brain frailty, and brain age offer new perspectives for assessing brain health reserve and predicting outcomes. This review systematically outlines the definitions, assessment methods, and clinical relevance of brain structure damage, brain frailty, and brain age, and examines their interrelationships. A deeper understanding of these relationships may allow more precise interpretation of individual differences in post-stroke recovery, providing valuable insights for establishing a multidimensional prognostic prediction system and advancing precision medicine in stroke care.

Patients with advanced non-small cell lung cancer (NSCLC) exhibit considerable variability in response to immunotherapy. The use of biomarkers or direct prediction of treatment efficacy is of great significance for identifying populations that may benefit. This review elaborates and analyzes the research progress, challenges, and future research directions of AI models in predicting immunotherapy biomarkers, developing non-traditional biomarkers, and predicting treatment response in advanced NSCLC.

CT based radiomics can extract high throughput quantitative features from medical images, providing a noninvasive and reproducible decision support tool for lung cancer management. In prognostic prediction, radiomics can be used to assess patient survival, risk of distant metastasis, and lymph node metastasis status, thereby assisting in risk stratification. For therapeutic evaluation, this technique enables early prediction of patient response to conventional chemoradiotherapy, immunotherapy, and targeted therapy, helping to identify potential beneficiary populations. By integrating radiomics features with clinicopathological information, more precise individualized treatment strategies can be developed. This article reviews the latest research progress of radiomics in prognostic prediction and treatment response assessment in lung cancer.

Accurate prediction of local invasion and distant metastasis in pancreatic ductal adenocarcinoma (PDAC) is crucial for therapeutic decision-making and prognosis. Radiomics enables high-throughput extraction of quantitative features from medical images. When combined with machine learning algorithms, it can further explore tumor heterogeneity in PDAC and provide a new approach for the preoperative, noninvasive assessment of tumor invasion and metastasis. Machine learning-based radiomics has been applied to predict vascular invasion, perineural invasion, lymph node metastasis, and liver metastasis in PDAC. Its performance is often superior to that of conventional imaging evaluation methods and can assist clinicians in formulating personalized treatment strategies. This article reviews the research progress and challenges of machine learning-based radiomics in predicting local invasion and distant metastasis of PDAC.

The classification of endometrial cancer (EC) has evolved from the traditional pathological dual-classification model to a molecular classification system encompassing four subtypes: DNA polymerase epsilon (POLE)-mutated, microsatellite instability (MSI), p53-mutated, and no specific molecular profile (NSMP). This system has been incorporated into the latest International Federation of Gynecology and Obstetrics (FIGO) staging criteria. As the preferred imaging method for preoperative evaluation of EC, MRI, together with functional imaging, radiomics, and deep learning technologies, has shown great potential for the non-invasive prediction of EC molecular subtypes. This article reviews the applications and research progress of MRI in the EC molecular classification, with the aim of providing a reference for precision medicine.

Quantitative MRI can detect early involvement of the synovium, cartilage, and bone in juvenile idiopathic arthritis (JIA) and further investigate its underlying pathophysiological processes, compared to conventional imaging examinations. Current quantitative MRI techniques include dynamic contrast-enhanced (DCE)-MRI, diffusion-weighted imaging (DWI), T₂ mapping, T_1ρ mapping and other data acquisition methods, which can provide preliminary evidence for early diagnosis, treatment planning, and evaluation of therapeutic response in JIA. This article reviews the advances in the application of quantitative MRI in JIA.

Osteoporotic vertebral fracture (OVF) is characterized by high incidence and high disability rate. Machine learning (ML)-based radiomics studies have increasingly focused on the full-process management of this disease. This paper describes the current research status of ML in three key aspects of OVF: prediction and risk assessment, diagnosis and classification/grading, and prognosis and treatment response evaluation. It highlights recent research progress of ML in solving core clinical problems such as OVF prediction, differentiation of acute and chronic OVF, and OVF refracture, aiming to provide a reference for more accurate and effective integration of ML technologies into the clinical diagnosis and treatment pathway of OVF.

Objective To analyze the CT and MRI manifestations of anterior mediastinal T-lymphoblastic lymphoma/acute lymphoblastic leukemia (T-LBL/ALL) in children. Methods The clinical data of 23 children with pathologically confirmed anterior mediastinal T-LBL/ALL were retrospectively collected. All patients underwent plain and contrast-enhanced CT scans prior to chemotherapy, and 4 cases additionally underwent plain and contrast-enhanced MRI examinations. Independent-sample t-test and Fisher’s exact test were employed to compare CT findings between T-LBL and T-ALL. The CT manifestations of all patients and the MRI features of the 4 cases were analyzed. Results All of 23 tumors were in mass form (100%) and had a large volume. Among them, 22 cases (95.7%) had well-defined margins, 7 cases (30.4%) were uniformly solid in density, and 16 cases (69.6%) exhibited cystic-solid components. After contrast enhancement, the solid components of the 23 tumors showed mild-to-moderate progressive enhancement with small intratumoral vessels traversing the lesions. The majority of patients showed marked mass effect on imaging. Most had compression of mediastinal large vessels (20 cases, 87.0%) and airways compression (19 cases, 82.6%). Pericardial and pleural effusions (20 cases, 87.0%) and pleural thickening (13 cases, 56.5%) were observed. A small number of patients developed venous thrombosis (1 case, 4.3%) and renal metastasis (8 cases, 34.8%). Cystic degeneration and necrosis were more likely to occur in T-LBL than in T-ALL (P=0.027). Conclusion The CT and MRI imaging manifestations of T-LBL/ALL in the anterior mediastinum of children have certain characteristics, and imaging follow-up can effectively evaluate the treatment effect and the progression of complications.