Ahmed E. Othman

Purpose: To evaluate the diagnostic confidence and image quality of deep-learning-enhanced ultra-high-resolution CT venography (CTV) in venous neurovascular imaging, compared with hybrid iterative reconstruction of ultra-high-resolution CT datasets and normal-resolution CTV. Methods: This retrospective, single-center study included 31 consecutive patients with suspected cerebral venous and dural sinus thrombosis who underwent CTV. CTV datasets were reconstructed using hybrid iterative reconstruction with (i) Normal-Resolution CTV (NR-CTV), (ii) an Ultra-High-Resolution CTV (UHR-CTV), and (iii) a deep learning algorithm specifically trained for neurovascular imaging, applied to the UHR-CTV datasets (DL-UHR-CTV). Quantitative analyses included signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and frequency power spectra (FPS). Qualitative evaluation of image quality, contrast, noise, artifacts, diagnostic confidence, and assessability of dural sinuses and intracranial veins was performed by three radiologists using a 4-point Likert scale. A descriptive analysis assessed concordant detection of venous pathologies across reconstructions. Results: SNR and CNR in DL-UHR-CTV were significantly higher than in UHR-CTV (SNR: p = 0.02; CNR: p = 0.01). NR-CTV and DL-UHR-CTV were similar. Compared to NR-CTV and UHR-CTV, DL-UHR-CTV showed a significantly lower FPS amplitude across wide frequency ranges. In qualitative analysis, DL-UHR-CTV outperformed UHR-CTV in all parameters ( p < 0.001) except vessel contrast ( p = 0.09) and was better than NR-CTV in all categories ( p < 0.001). All 20 venous pathologies (12 patients) were detected on all three reconstructions without false-positive findings. Conclusion: Deep-learning-based reconstruction significantly enhances image quality and venous visualization in ultra-high-resolution CT venography compared with conventional iterative reconstruction, enabling detailed depiction of small venous structures.

Purpose: Deep learning reconstruction may enable substantial radiation dose reduction in ultra–high-resolution (UHR) temporal bone CT, but its performance across clinically relevant dose levels remains insufficiently defined. This study evaluated vendor-specific deep learning reconstruction (DLR) compared with hybrid iterative reconstruction (HIR) in UHR volume-mode temporal bone CT and aimed to propose an indication-adapted framework for dose optimization. Methods: In this single-center body donor study, 20 temporal bones from 10 donors without pathologic findings were scanned at 11 dose levels (CTDI _vol , 5.1–30.7 mGy). Images were reconstructed using HIR at standard resolution (0.5 mm, 512 ² matrix) and UHR (0.25 mm, 1024 ² and 2048 ² matrices), as well as DLR (AiCE Inner Ear) at UHR (0.25 mm, 1024 ² matrix). Three radiologists independently assessed 10 anatomical structures using a 5-point Likert scale. Signal-to-noise ratio was measured in predefined regions. Exploratory structure-specific generalized estimating equation models were used to estimate the probability of achieving optimal image quality as a function of radiation dose and reconstruction technique. Results: DLR significantly improved visualization of osseous structures compared with HIR ( p < 0.001). The exploratory dose-response analysis demonstrated reconstruction-specific trajectories, with DLR achieving higher predicted probabilities of optimal image quality than HIR at lower dose levels. For routine clinical indications, diagnostically adequate image quality was achieved with DLR at a CTDI _vol of 13.3 mGy, representing a 48% dose reduction compared with HIR (CTDI _vol 25.6 mGy). For high-detail assessment of delicate microanatomical structures, including the stapes and cochlea, DLR achieved excellent visualization at a CTDI _vol of 17.4 mGy, corresponding to a 43% dose reduction compared with HIR (CTDI _vol 30.7 mGy). Conclusion: Vendor-specific DLR enables substantial radiation dose reduction in UHR volume-mode temporal bone CT while maintaining high diagnostic image quality. These findings support an indication-adapted framework for dose optimization, suggesting a routine clinical protocol at CTDI _vol 13.3 mGy and a microanatomy-optimized protocol at CTDI _vol 17.4 mGy.

Background: Single-shot echo-planar imaging DWI (ss-EPI DWI) is susceptible to geometric distortions near air–tissue interfaces, limiting diagnostic accuracy. Multishot readout-segmented DWI (RESOLVE DWI) mitigates these artifacts but requires longer acquisition times. This study evaluated whether ss-EPI DWI with a deep learning–enabled correction of static field inhomogeneities (DL DWI) can achieve effective distortion correction while preserving diagnostic reliability and reducing scan time. Objectives: To compare image quality, geometric distortion, lesion detectability, and diagnostic confidence of ss-EPI DWI with DL-enabled correction of static field inhomogeneities (DL DWI) against conventional ss-EPI DWI (without correction) and RESOLVE DWI. Materials and Methods: In this prospective single-center study (April 2025 to August 2025), 100 patients undergoing clinically indicated brain MRI were enrolled following written informed consent. All patients underwent DL ss-EPI DWI with DL correction (1 min 29 s) and RESOLVE DWI (3 min 17 s); ss-EPI DWI was reconstructed by disabling DL correction for all patients. Qualitative and quantitative analyses were performed. Three board-certified radiologists rated image quality, lesion detectability, distortion, and diagnostic confidence on a 5-point Likert Scale. Two quantitative measures of distortion were applied: geometric conformity with anatomic 3D T1-weighted data through coregistration methods and gray-value distributions within volumes of interest in susceptibility-affected regions (temporal lobe, cerebellum, and brainstem). Results: DL DWI significantly outperformed ss-EPI DWI across all qualitative parameters, including diagnostic confidence (median, 4.0 vs. 3.0; P < 0.001), overall image quality (median, 4.0 vs. 3.0; P < 0.001), and geometric distortion (median, 4.0 vs. 3.0; P < 0.001). RESOLVE DWI achieved the highest ratings overall (median, 5.0). Lesion detectability was comparable between DL DWI and RESOLVE DWI (median, 4.0 for both; P > 0.08), while both significantly outperformed ss-EPI DWI ( P < 0.001). Quantitative analysis confirmed superior geometric accuracy of DL DWI over ss-EPI DWI, with lower coregistration cost-function values and normalization of gray-value distributions in susceptibility-affected regions. Conclusion: DL DWI improved image quality and geometric accuracy compared with uncorrected ss-EPI DWI. Subjective lesion-detectability ratings were similar between DL DWI and RESOLVE DWI, while DL DWI required less than half the acquisition time. These findings support DL-based distortion correction as a promising, time-efficient approach for brain diffusion-weighted imaging that warrants confirmation in a predefined noninferiority design.

Background and ObjectivesCardiac abnormalities are among the most common and complex risk factors of pediatric acute ischemic stroke (AIS). Although endovascular treatment (EVT) is increasingly use...

Objectives: Vessel wall magnetic resonance imaging is important in the diagnosis of intracranial vascular diseases. Until now, adequate clinical implementation has required the use of 3 T scanners, which limits wider availability. This study aims to evaluate the interchangeability of a 3D turbo spin echo sequence with variable flip angle MRI sequence (T1 3D Sampling Perfection with Application optimized Contrast using different flip angle Evolution, SPACE) with deep learning image reconstruction for a 1.5 T scanner, with an established sequence (T1 3D SPACE) for a 3 T scanner. In addition, image quality parameters and diagnostic confidence are to be compared. Materials and Methods: Patients were prospectively examined using a T1-weighted turbo spin echo with variable flip angle at 3 T and a similar newly developed sequence with deep learning-based image reconstruction at 1.5 T. The images were read by 3 experienced radiologists regarding the presence of pathologic enhancement of the vessel wall for the interchangeability analysis. Image quality and diagnostic confidence were evaluated using 5-point Likert scales. Interreader and intrareader reliability were calculated for all parameters. Results: A total of 50 participants (42.6 ± 21.2 y, 26 males) were included. Regarding the presence of pathologic vessel wall enhancement, there was a high interreader agreement [Fleiss kappa = 0.72 (95% CI: 0.62; 0.81) for 1.5T, 0.78 (0.69; 0.88) for 3 T]. The sequences at 1.5 T and 3 T were interchangeable (IEI = 0.008). Image quality at 1.5 T was diagnostic in all cases, but rated superior at 3 T ( P < 0.001). Diagnostic confidence was also slightly but significantly higher at 3 T ( P < 0.001). Conclusions: The utilization of deep learning reconstruction enables vessel wall imaging in 1.5 T MRI. This new MRI sequence was interchangeable with established sequences at 3 T for the detection of pathologic vessel wall enhancement. This development could significantly improve the availability of vessel wall imaging in the clinical routine.

Background: Manual lesion measurements remain the standard for assessing oncologic treatment response, despite being time-consuming and prone to substantial inter-reader variability, which may lead to inconsistent response classification and subsequent variability in treatment decisions. Purpose: To evaluate the impact of an artificial intelligence (AI) system for assisted lesion measurement on reading time and measurement consistency in follow-up CT examinations using the Response Evaluation Criteria in Solid Tumors (RECIST 1.1). Methods and Materials: In this retrospective reader study, follow-up chest-abdomen-pelvis CT examinations from 212 oncology patients collected at two Dutch hospitals were assessed by 23 readers (15 radiologists, 8 residents) recruited from 11 international institutions under three conditions: unassisted, AI-assisted, and expert-assisted (using a prior radiologist’s unassisted measurement). To prevent bias related to the source of the measurements, readers were informed that all support was AI-derived. Primary outcomes were reading time to completion and inter-observer measurement variability. For each outcome, a Bayesian generalized linear mixed model was used to analyze the results. Results: AI assistance significantly reduced per-patient reading time versus unassisted reading (-36.0 s; 95% CI: -53.0, -22.1). At the lesion level, it was associated with a small increase in variability relative to the expert-derived reference standard (1.32 mm; 95% CI: 0.83, 1.91). At the patient level, AI assistance did not meaningfully affect change in the sum of longest diameters (SLD; -0.38 mm; 95% CI: -1.57, 0.72), but increased RECIST outcome agreement by 7.7% (95% CI: 2.8, 12.7) compared with unassisted reading. Expert-assisted reading yielded even higher inter-reader agreement (13.3%; 95% CI: 8.6, 18.1). Conclusion: AI assistance reduced reading time and improved patient-level RECIST agreement, despite a small increase in lesion-level measurement variability. These findings suggest that AI-assisted RECIST assessment may improve workflow and response classification consistency, while also providing a benchmark from expert-assisted reading for future AI development.

Purpose: Digital Subtraction Angiography (DSA) is an X-ray-based imaging modality intimately related to minimally invasive procedures in interventional radiology, cardiology, vascular and neurologic surgery. Emulating tomographic methods like 3D vessel reconstruction and flat-panel detector CT perfusion imaging increases its diagnostic utility. This study demonstrates a hardware and software setup expanding DSA capability to functional perfusion imaging by assessing planar cerebral perfusion at runtime. Methods: The setup uses an HDMI video splitter and frame-grabber for duplicating the video output of the intervention suite to an arbitrary machine at the time of image acquisition. A selection of methods, including numerical approximation and neural network inference of fitted curve features, is applied to create perfusion parameter maps and compare their respective completion times over a set of angiographic runs. Results: We identified two distinct perfusion estimation methods able to yield results within 1–2 s, signal deconvolution using single-value decomposition (SVD) and numerical curve feature estimation, as well as a variation on conventional curve fitting that massively shortened calculation times from five minutes to clinically feasible 30 s. Conclusion: Directly accessing image data outside of the angiography suite enables real-time angiographic cerebral perfusion evaluation in the clinical workflow. This way, on-the-fly analysis of angiograms in clinical settings, e.g., angiographic perfusion, is made possible, facilitating future clinical studies.

Purpose: As MRI frequently requires general anesthesia in pediatric patients, there is an undersupply in clinical routine. This may result in delayed examinations or use of CT, especially in emergency settings. To enable ad-hoc MRI scans and rule out increased intracranial pressure or intracranial mass lesions, we combined various acceleration techniques with Deep-Learning reconstruction, generating ultra-fast diagnostic images sufficient to exclude critical pathologies. Methods: Thirty-six MRI datasets of infants with a median age of 35.2 months (SD ± 23.2) and anesthesia-free imaging were retrospectively evaluated. Imaging was performed using ultra-fast T2-weighted sequences in three planes (slice thickness 5 mm; total acquisition time 47s). Four readers evaluated subjective image quality using a 5-point Likert-scale. Readers were asked to indicate how safe they felt about assessing possible midline displacement or mass lesion. For semi-quantitative analysis, readers reported diameters of lateral and third ventricles. Gwet’s AC2 and intraclass correlation (ICC) were used for interrater agreement. Results: 94.4% of datasets showed at least acceptable diagnostic confidence. Readers felt confident excluding acute intracranial pathology. 52.1% of ultra-fast sequences demonstrated good to excellent image quality. 72.6% were rated with good or excellent diagnostic confidence. Interrater reliability demonstrated almost excellent agreement of diagnostic confidence (Gwet’s AC2 ≥ 0.886) and image quality (Gwet’s AC2 ≥ 0.942). Excellent agreement regarding ventricular width (ICC values ≥ 0.966) was shown for all measurements. Conclusion: The use of deep-learning reconstruction algorithms in pediatric brain MRI is feasible, allowing anesthesia-free emergency imaging. This will reduce periprocedural risk, number of necessary CT scans, and lower healthcare costs.

Rationale and Objectives: Diffusion-weighted imaging (DWI) of the head and neck is essential for various clinical applications but is often hampered by artifacts and reduced image quality. Deep learning (DL) reconstruction has the potential to enhance the quality of head and neck DWI. This study aims to evaluate the performance of an accelerated, DL-reconstructed DWI (DWI _DL ) in terms of image quality and diagnostic confidence. Materials and Methods: This retrospective study included patients who underwent clinically indicated head and neck DWI at 1.5 T and 3 T between August 2023 and January 2024 at a tertiary care center. Imaging was performed at low b‑values (0 or 50 sec/mm ² ) and high b‑values (800 sec/mm ² ), and apparent diffusion coefficient (ADC) maps were computed. After acquiring standard single-shot echoplanar imaging DWI sequences, the raw MR datasets underwent simulated acceleration by reducing the number of signal averages. These accelerated exams were then reconstructed using a novel DL-based algorithm that combined DL-based k‑space to image reconstruction with DL-based super-resolution processing (DWI _DL ). Three readers analyzed the images using a visual Likert score to evaluate image sharpness, artifacts, noise, overall image quality, and diagnostic confidence. Comparisons were made using the Wilcoxon signed-rank test. A quantitative analysis of signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and apparent diffusion coefficient values (ADC) was also performed. Results: The study included 30 patients (mean age, 55 ± 19 years; range, 24–84; 18 men) with various pathologies. Scan times were reduced by 67% at 1.5 T and up to 55% at 3 T. The quantitative analysis revealed a minimal but statistically significant decrease in SNR and CNR in the deep learning-reconstructed images ( p = 0.002 and p < 0.001, respectively). However, readers reported no significant differences between DWI and DWI _DL regarding image quality parameters or diagnostic confidence for both low and high b‑value images, as well as the ADC (all p > 0.05). Conclusion: DL reconstruction of head and neck DWI is feasible, significantly reducing examination time without compromising image quality or diagnostic confidence. This technique enables accelerated and effective diagnostic DWI of the head and neck.