Computed tomography (CT)-guided interventions are central to interventional radiology (IR) but remain highly operator-dependent. Robotic assistance has been proposed to improve needle accuracy, reproducibility and radiation safety. While robotics are well established in surgery, their role in CT-guided IR remains evolving. This review synthesises current evidence, maps robotic CT-guided interventions onto the IDEAL framework for procedural innovation, and evaluates where these systems provide measurable clinical value. A narrative review was conducted according to established recommendations for high-quality narrative reviews and informed by SANRA principles. PubMed, Scopus and clinical trial registries were searched for studies evaluating robotic CT-guided percutaneous interventions. Evidence from preclinical, feasibility and comparative clinical studies was synthesised qualitatively and interpreted using the IDEAL framework. Preclinical studies consistently demonstrate improved targeting accuracy and fewer needle adjustments compared with freehand techniques. Clinical studies and systematic reviews support feasibility and safety across biopsy and ablation applications, with current clinical evidence concentrated in the liver, lung, kidney, and musculoskeletal system, particularly for long, oblique, and out-of-plane trajectories. Operator radiation exposure is frequently reduced, while procedure time and patient dose effects are variable. Most published studies correspond to IDEAL stages 1–2b, with emerging Stage 3 randomised evaluation; long-term and registry data remain limited. Robot-assisted CT-guided interventions represent precision-enabling technologies but are not yet standard of care. Current evidence supports selective use in technically complex procedures. Wider adoption will depend on robust comparative studies, economic evaluation and structured implementation frameworks. Question What evidence supports robot-assisted CT-guided interventions in IR, and in which procedures do robotic systems provide clinical benefit? Findings Current clinical evidence is concentrated in liver, lung, kidney, and musculoskeletal interventions, demonstrating improved targeting precision and reduced needle adjustments and operator radiation exposure. Clinical relevance Robot-assisted CT-guided interventions may improve procedural precision and reduce radiation exposure in complex interventions, but wider adoption depends on stronger comparative evidence demonstrating meaningful clinical benefit, workflow efficiency, and cost-effectiveness.
This study provides a Rapid Analysis and Processing of Image Data (RAPID) framework that combines deep learning-based CT topogram analysis with DICOM spatial geometry to enable reliable anatomical labelling of CT series independent of inconsistent textual metadata. In this single-centre retrospective study, three YOLOv8-based models comprising the RAPID framework were trained on CT topograms to perform global anatomical classification, body-region detection, and landmark detection. Classification used 83207 topograms (20,802 test), while landmark and body region detection models were trained on 2000 (500 test) and 1926 (481 test) topograms, respectively, collected between 2003 and 2022. Model performance was evaluated using the F1 score and mAP50, with additional external validation on the external cohort. Furthermore, three radiologists independently reviewed 150 randomly selected predictions for detection models using a Likert-scale-based clinical assessment with inter-rater agreement. Across a total of 65,250 patients (median age, 62 years; interquartile range, 23; 44
To identify clinically established and frequently utilized routine head computed tomography (CT) acquisition techniques in children and adults that achieve lower radiation. The level of the analysis is at the CT scan (series) level. CT acquisition parameters were analyzed from a large, international CT dose registry. Scans were analyzed separately in adults and children, and by approach (helical or axial), resulting in four strata. K-means clustering, an unsupervised machine learning approach, assigned the scans into clusters based on the following acquisition parameters: Tube current, voltage, collimation, scan length, and for helical scans, pitch. Effective patient diameter, unadjusted and patient size-adjusted dose-length product (DLP), and volume CT dose index (CTDIvol) were compared across clusters. 864,182 CT scans in adults and 40,027 scans in children between January 1, 2015, and March 11, 2021, were included in the analysis. Scans were classified into seven helical and six axial clusters in adults, six helical and four axial clusters in children. In all four strata, the mean size-adjusted DLP varied more than two-fold, and up to 3.4-fold, across the clusters. For example, for adult helical scans (the largest strata), the size-adjusted DLP ranged from 372 mGy·cm to 998 mGy·cm (relative dose 2.7) across clusters. Clusters with the lowest radiation generally utilized lower tube current and voltage, but protocols utilized various approaches to result in lower doses. There remains a large variation in the radiation dose within frequently used head CT acquisition techniques. Clustering analysis enables identifying protocols with lower doses. Question Is there a large variation in the doses used for routine head CT, and can we identify best-practice, lowest dose approaches for scanning? Findings Across each stratum (adult/pediatric, helical/axial), the mean size-adjusted DLP varied by more than two-fold across the CT scan clusters, reflecting significant differences in radiation. Clinical relevance Many routine head CT acquisition techniques use excessive radiation doses. Standardizing practice to the lowest dose clusters would result in significant reductions in patient dose.
Artificial intelligence (AI) has become an increasingly prominent force in medicine, driven by rapid technical advances and a growing number of clinical applications. As the field matures, it now increasingly moves from experimental development toward a phase of broader implementation. However, debates surrounding medical AI are often shaped by exaggerated risks and overly optimistic expectations. This paper seeks to contribute to a more balanced and realistic discussion. Rather than framing AI as either an existential threat or a universal solution, we advocate for an open-minded, evidence-based understanding of AI as a tool to support healthcare and discuss current and emerging challenges related to clinical validation, human–AI interaction, bias and discrimination, education, agentic AI, and the development and maintenance of trust. Question Discussions about AI in medicine continue to be dominated by exaggerated risks and overly optimistic expectations. Findings We provide a more realistic evaluation of the current opportunities of AI in medicine and want to highlight some genuine challenges that lie ahead. Clinical relevance AI is here to stay in medicine. Focusing on real and present challenges, rather than being distracted by exaggerated risks, as well as responsible expectation management, will be key to its success.
Accurate longitudinal nodule matching is a critical technical prerequisite for automated growth rate (volume doubling time) assessment in lung cancer screening. This study evaluated an artificial intelligence (AI) pulmonary nodule analysis system in all 361 UK Lung Cancer Screening (UKLS) trial participants who underwent a 3-month follow-up low-dose computed tomography (LDCT) scan. The pulmonary AI independently evaluated these baseline scans using an updated volume threshold (solid component ≥ 100 mm³ per NELSON 2.0/EUPS protocol) for cases requiring 3-month follow-up. To assess true algorithmic robustness, all AI-detected baseline candidate nodules (≥ 100 mm³) proceeded to fully automated longitudinal matching without any manual selection. The pulmonary AI identified 181 participants with 378 baseline nodules ≥ 100 mm³. In total, 39 nodules had naturally resolved at follow-up. The pulmonary AI achieved an 83.5
O-RADS MRI improves differentiation of benign adnexal lesions managed by a gynecologist and malignant lesions managed by a gynecologic oncologist. To compare O-RADS MRI visual assessment (VA) of mass enhancement with a semi-quantitative intensity curve of enhancement (SIC), a modified manual version of the time intensity curve. We conducted a retrospective MRI review of patients who underwent surgical resection between January 2008 and December 2018 at two referral centers. O-RADS MRI score was assigned by two radiologists with different experience. O-RADS score ≥ 4 was considered suspicious for malignancy. Per-lesion diagnostic accuracy and impact on surgical management were compared between VA and SIC in lesions with solid components. Inter/intra-reader agreement was assessed. Borderline and invasive lesions were considered malignant. Three hundred forty-one adnexal masses in 321 women (ages 18–89) were assessed; 35.2
To evaluate the spatial heterogeneity of viscoelastic properties in pancreatic cancer, to correlate MR elastography (MRE) parameters with clinicopathological factors indicating tumor aggressiveness, and to identify risk factors for patient survival. This retrospective study included 96 treatment-naive pancreatic cancer patients undergoing multifrequency MRE between September 2018 and May 2024. High-resolution parametric maps were generated to measure stiffness and fluidity in central, peripheral, and whole-tumor regions. Correlations between MRE parameters and tumor stage, grade, vascular and perineural invasion, regional lymphadenopathy, and distant metastasis were analyzed. Kaplan–Meier and Cox proportional hazards models were used to identify prognostic factors in patients with and without R0 resection, respectively. Peripheral fluidity of the tumor was higher than central fluidity (1.23 ± 0.25 vs 0.95 ± 0.19 rad, p < 0.001), whereas stiffness showed no zonal difference (p = 0.07). Tumor stiffness correlated positively with vascular invasion (ρ = 0.21, p = 0.04). Peripheral tumor fluidity correlated positively with tumor stage (ρ = 0.28, p = 0.007), perineural invasion (ρ = 0.31, p = 0.02), regional lymphadenopathy (ρ = 0.29, p = 0.005), and distant metastasis (ρ = 0.26, p = 0.01). In 48 R0-resected patients, higher peripheral fluidity was associated with shorter disease-free survival (DFS) (9.1 vs 19.4 months, p = 0.009) and was identified as an independent predictor of shorter DFS (hazard ratio, 2.15; 95
Current hepatocellular carcinoma (HCC) surveillance relies on semiannual ultrasonography, which has limited sensitivity for early-stage detection. We evaluated the effectiveness of HCC surveillance using gadoxetic acid-enhanced MRI at risk-stratified screening intervals in patients with cirrhosis. We prospectively enrolled 86 patients with cirrhosis between May 2021 and October 2022. All patients underwent HCC surveillance using gadoxetic acid-enhanced MRI for 3 years at risk-stratified screening intervals ranging from 3 to 36 months, according to the assigned risk category defined by LI-RADS criteria on baseline MRI (low, intermediate, and high risk with LI-RADS LR-3 or LR-4 observations). The primary outcome was surveillance failure, defined as the detection of HCC beyond the Milan criteria. HCC was diagnosed in 13 patients during the study period. The HCC incidence differed markedly across the MRI risk categories, ranging from 1.9 per 100 person-years in patients with low risk to 73.4 per 100 person-years in those with LR-4 observations. The primary endpoint was met with no surveillance failures (0
The sensitivity of mammographic screening is lower for women with mammographically dense vs fatty breasts. We aimed to explore automated mammographic breast density and malignancy risk scores generated by an artificial intelligence (AI) model for breast cancer detection, stratified by mammography vendor. This retrospective study included information from 200,000 examinations within BreastScreen Norway. An automated volumetric breast density (VBD) assessment and risk scores were obtained from a commercial AI model. The continuous VBD output from the AI model was categorized into four groups, VBD 1-4, with 15
To analyze the factors for predicting 1-year outcomes and optimize surgical planning for Meso-Rex bypass (MRB) in children with extrahepatic portal vein obstruction (EHPVO). From October 2014 to July 2024, children with EHPVO after MRB were retrospectively analyzed. The number of MRB, the type of bypass vessel, and the position of the lower anastomosis were collected. Predictive variables include the diameter of the connection between the left and right portal veins, the Rex vein (RV) diameter, the number of RV branches, the diameter of the thickest RV branch, and the bypass-to-superior mesenteric vein (SMV) angle, which were measured by Contrast-enhanced CT (CECT) one week after MRB. The upper and lower anastomotic stenosis were diagnosed by Ultrasound and CECT one year after MRB. A total of 153 children (median age 72 months, 87 males) were included in this study. Univariate factors include the number of MRBs, the type of bypass vessel, and CECT variables (except the diameter of the thickest RV branch). Numbers of MRB and RV branches were retained in the final model for upper anastomotic stenosis. The combination of MRB > 1 and RV branches ≤ 4 predicts it post-MRB with 75
To compare reading time, cancer detection rate (CDR), and abnormal interpretation rate (AIR) between AI-assisted and non-AI-assisted periods in screening and diagnostic mammography performed in routine clinical practice. We prospectively collected reading times for consecutive two-view full-field digital mammography interpreted by four radiologists between August 2023 and July 2024. Both screening and diagnostic examinations were included. A commercially available AI system was integrated into the clinical workflow, with results displayed or hidden on a monthly basis. Reading time, CDR, and AIR were compared between two periods. For reading time analysis, a subset of 2917 examinations with times ≤ 5 min was included to minimize the impact of non-interpretive interruptions. Reading time was extracted from the PACS log. Among 4577 mammography examinations (mean age 51.7 ± 10.4 years), the overall CDR was higher during the AI-assisted period (22.3 vs 11.5 per 1000; p = 0.005). AIR did not differ for screening mammography (9.5
