Background Coronary artery disease is a leading cause of cardiovascular death. Conventional noncontrastu2010enhanced 3.0 T coronary magnetic resonance angiography (CMRA) is safe but limited by long scan times and motion artifacts. Compressed sensing (CS) can accelerate imaging, yet the optimal acceleration factor remains undefined. This study aimed to evaluate the diagnostic accuracy of 4.5u2010fold CS (4.5 u00D7 CS)u2010accelerated 3.0 T noncontrastu2010enhanced CMRA for detecting coronary artery stenosis, using invasive coronary angiography as the reference standard. Methods Thirty patients with suspected coronary artery disease who underwent CMRA, coronary computed tomography angiography, and invasive coronary angiography at the Affiliated Hospital of Guizhou Medical University from December 2022 to November 2024 were enrolled. The diagnostic performance of CMRA for detecting u2265 50.00% stenosis was evaluated using the kappa test (agreement), Wilcoxon signedu2010rank test (continuous variables), and calculations of sensitivity, specificity, and accuracy. Results The 4.5 u00D7 CS protocol shortened CMRA scan time by 52.30% (to [7.2 u00B1 1.3] min) while maintaining image quality (signalu2010tou2010noise ratio, 25.3u201330.2; contrastu2010tou2010noise ratio, 19.6u201324.5). CMRA showed excellent agreement with invasive coronary angiography (kappa = 0.905, p u0026lt; 0.001), with strong diagnostic performance at the segment level (sensitivity 86.90%, specificity 99.01%, accuracy 95.45%), the vessel level (sensitivity 94.59%, accuracy 79.17%), and the patient level (sensitivity 92.00%, accuracy 90.00%). Conclusions 4.5 u00D7 CSu2010accelerated 3.0 T noncontrastu2010enhanced CMRA achieves a favorable balance between efficiency, image quality, and diagnostic accuracy. As a radiationu2010 and contrastu2010free modality, it may be particularly valuable for coronary artery disease screening and followu2010up (e.g., in contrastu2010allergic or renally impaired patients) and shows strong potential for clinical translation.
Background In clinical settings, particularly in radiology, artificial intelligence and large language models have already demonstrated promising results. They have the potential to become an integral component of radiologists' workflows in the future, and neuroradiology is likely to benefit from their increasing integration into clinical practice. This study aims to evaluate Chat Generative Preu2010trained Transformer 4 Omni (ChatGPTu20104o) as a supportive tool in neuroradiology by assessing its ability to generate accurate radiology reports from imaging and patient history and its impact on radiologists' diagnostic accuracy. Methods Retrospective analysis was conducted using only radiological images and brief patient admission history from 30 neuroradiology cases published publicly between 2015 and 2024. ChatGPTu20104o generated radiology reports with final and differential diagnoses. One radiologist and one radiology resident independently reviewed cases without and with the reports created by ChatGPTu20104o at a 4u2010week interval. Diagnostic accuracy was compared to the published goldu2010standard histopathological diagnoses. Also, ChatGPTu20104o was asked to detect the orientation, sequence, and contrast usage. To compare the accuracies, exact McNemar's test and Wilcoxon signedu2010rank test with Bonferroni correction was used for statistical analysis. Results Evaluation of radiology reports demonstrated an accuracy of final diagnosis from ChatGPTu20104o 33.33% (10/30), experienced radiologist 56.67% (17/30), and radiology resident 23.33% (7/30) (p u0026lt; 0.039 for experienced radiologist vs. ChatGPTu20104o and p u0026lt; 0.375 for resident vs. ChatGPTu20104o). Accuracy rates for differential diagnosis were: ChatGPTu20104o 60.50% u00B1 21.30%, experienced radiologist 64.83% u00B1 17.24%, and radiology resident 42.50% u00B1 16.90% (p = 0.242 for experienced radiologist vs. ChatGPTu20104o and p = 0.0004 for resident vs. ChatGPTu20104o). Diagnostic performance with the use of ChatGPTu20104o as a support tool demonstrated an accuracy of final diagnosis: experienced radiologist 56.67% (17/30) and radiology resident 30.00% (9/30) (p = 1.000 and p = 0.500, respectively) and of differential diagnosis: experienced radiologist 66.50% u00B1 19.30% and radiology resident 49.50% u00B1 15.56% (p = 0.500 and p = 0.013, respectively). Conclusions ChatGPTu20104o demonstrated limited autonomous and supportive diagnostic accuracy in complex neuroradiology cases. While its assistance did not improve the experienced radiologist's performance and produced only a modest, nonu2010significant improvement for the resident, a trend toward improved differential diagnosis generation was observed. These findings suggest that multimodal large language models may have limited supportive value, particularly for less experienced clinicians, and require validation in larger studies.
ABSTRACT Background Self‐supervised learning (SSL) has emerged as a promising approach in medical image analysis, offering the ability to learn robust feature representations from unlabeled data. However, the extent to which these embeddings align with clinically interpretable features remains largely unaddressed. This study aims to assess the degree to which SSL embeddings encode radiomics‐aligned, interpretable features within multiphase renal computed tomography (CT). Methods We analyzed four‐phase contrast‐enhanced CT scans of renal tumors, including the noncontrast, corticomedullary, nephrographic, and excretory phases. Radiomic features included first‐order statistics, texture‐based features, and shape descriptors. SSL embeddings were obtained from three models: Simple framework for contrastive learning of visual representations (SimCLR), distillation with no labels (DINO)—a self‐distillation with no labels method, and bootstrap your own latent (BYOL). Feature‐level alignment was quantified using pairwise Spearman correlation analyses between SSL embeddings and radiomic features. Results DINO demonstrated the highest alignment with radiomic features, particularly texture‐based features, such as gray‐level size zone matrix (GLSZM) (up to 22%) and neighboring gray tone difference matrix (NGTDM) (up to 35%), followed by SimCLR. BYOL showed minimal alignment across all feature families. Among imaging phases, the nephrographic phase exhibited the strongest correlations overall. Conclusions These findings indicate that SSL models, particularly DINO and to a lesser extent SimCLR, are capable of encoding patterns that align with established radiomic descriptors, especially texture‐based features. This partial alignment underscores the potential of incorporating radiomics‐informed metrics into the evaluation framework for SSL models, enabling more interpretable and clinically relevant model selection in medical imaging applications.
Background Sacral terminal filar cysts often cause pain, numbness, urinary/bowel dysfunction, and sexual dysfunction due to cerebrospinal fluid (CSF) leakage at the filum terminaleu2013dural sac junction. Accurate localization of the fistula is critical for surgical success because misidentification may lead to persistent symptoms and a requirement for reoperation. Although magnetic resonance imaging (MRI) and computed tomography (CT) are standard for diagnosis, manual localization of the fistula relies heavily on expert experience, limiting reproducibility. This study introduces an artificial intelligence (AI)u2010driven approach to overcome these limitations. Methods This study developed a twou2010stage AI algorithm using the Qwen large model for MRIu2013CT pattern matching. First, MRI slices were classified into three anatomical patterns (subarachnoid space, transitional, or cyst level) based on vertebral canal morphology, with the highestu2010scoring pattern taken to indicate the fistula level. Second, corresponding CT slices were matched using shape features of the vertebral canal. The algorithm was validated on three typical cases from our department, with the AIu2010generated localizations being compared with expert manual annotations. Results The AIu2010identified MRI slices (slices 12, 6, and 7 for cases 1u20133, respectively) matched expertu2010annotated levels. The corresponding AIu2010identified CT slices for cases 1 and 2 (slices 146 and 183) aligned perfectly with the manual records, whereas that for case 3 (239) differed by only two slices (1.4 mm) from the manual one (237). Threeu2010dimensional rendering of the CT confirmed the fistula's transitional location between the dural sac and cyst, validating anatomical accuracy. The algorithm demonstrated high robustness in crossu2010modal matching. Conclusions This study presents an AIu2010based approach for locating the fistula in sacral terminal filar cysts, which shows potential for reducing expert reliance and enhancing accessibility. However, the small sample size limits the generalizability of our preliminary results, which are primarily based on descriptive comparisons and lack rigorous statistical analysis.
