Academic Radiology最新文献

Rationale and objectives: To evaluate the clinical performance of a diffusion model-based motion correction algorithm for portable brain CT.

Materials and methods: We retrospectively collected 67 portable brain CT scans with corresponding fixed CT scans acquired within ±2 days as reference. A pre-trained diffusion model was applied to correct motion artifacts in the portable scans. Each case yielded three volumes as follows: original (motion group), corrected (corrected group), and fixed (reference group). Images were reviewed in randomized order by three professional readers (one neuroradiologist, one neuroradiology fellow, and one radiology resident), with at least two weeks between sessions to reduce recall bias. Eight lesion types and four image quality metrics were scored using a 5-point Likert scale. ACR phantom testing was performed to assess compliance with diagnostic image quality standards.

Results: Corrected images significantly outperformed motion images in all image quality metrics (improvement: 0.33-0.79, p<0.001), except for sharpness (p = 0.34). Diagnostic confidence improved from 2.52 to 2.86. Lesion detectability remained comparable before and after correction, with no significant differences in agreement rates (McNemar's p>0.10) or AUCs (DeLong's p>0.06) across all lesion types. Agreement rates ranged from 0.866 to 0.985 in the corrected group against the reference, and AUCs from 0.788 to 0.964. The net reclassification index was 2.66%. Corrected images passed all ACR criteria in phantom testing.

Conclusion: The diffusion model-based algorithm effectively improves image quality and diagnostic confidence without compromising lesion detection, supporting its potential for clinical use in portable brain CT.

Radiological reports are essential clinical documents often written in highly technical language that is challenging for patients to comprehend. Despite advancements in digital imaging and reporting technologies, the inherent complexity of radiology reports creates significant barriers to effective patient understanding. Recently, large language models (LLMs) have emerged as a promising solution to simplify radiological reports. Therefore, this narrative review aims to provide a comprehensive overview of LLMs for simplifying patient-centered radiology reports. We examined 19 studies evaluating various LLMs including GPT-3.5, GPT-4, Claude, Gemini, and others across multiple imaging modalities. All studies reported descriptive/consistent improvements in readability metrics, with simplified reports typically achieving 5th-8th grade reading levels compared to the original 10th-14th grade levels. However, many studies identified accuracy concerns, with reports containing a range of omissions, commissions, and distortions depending on modality and model. Building upon these findings, we discuss medicolegal considerations, workflow integration challenges, and strategies for effective LLM implementation. We also explore potential impacts on radiologist workflow, including the impact of LLM biases and liability for simplified reports. Despite promising results, significant challenges remain in ensuring accurate simplification across diverse patient populations while maintaining clinical precision. In conclusion, this review underscores the transformative potential of LLMs in enhancing patient understanding of radiological findings while highlighting the need for careful implementation with appropriate oversight mechanisms.

Rationale and objectives: Innovative, evidence-based, and feasible educational interventions to teach pediatric musculoskeletal (pMSK) radiograph interpretation to radiology post-graduate trainees (R-PGT) are currently lacking.

Purpose: We evaluated the effectiveness of a pMSK radiograph education intervention in improving the identification and risk stratification of fractures and dislocations. We also determined cases most at risk of diagnostic error.

Methods: This was a multicenter prospective cross-sectional study in a convenience sample of R-PGT practicing in the United States and Canada. The web-based education intervention included 1609 pMSK extremity radiographs organized into six anatomic regions. R-PGT deliberately practiced identifying if there was a fracture/dislocation present or absent, and if present, they located and risk-stratified the fracture. Participants completed cases until they achieved a performance standard.

Results: We enrolled 100 R-PGT and derived 48,166 unique case interpretations. From the initial to final 25 case completions, there were learning gains in diagnostic sensitivity (14.9%; 95% CI 13.4, 16.4), fracture location accuracy (14.1%; 95% 12.6, 15.5), and risk stratification (23.6%; 95% CI 21.5, 25.7). Of the 100 R-PGT, 77.5% (95% CI 71.1; 83.1) achieved the performance standard in at least one anatomic region in a median of 173 cases (IQR 94, 315) or a median of 41.5 min (IQR 22.6, 76.6). There was a higher odds of correctness in older versus younger children (OR=1.3; 95% 1.2, 1.4) and those without versus with a suspicion for non-accidental injury (OR=2.0; 95% CI 1.6, 2.4). The most frequent locations among the 171 high-risk false negative cases were the elbow (n=48 [28.1%]), pelvis (n=39 [22.8%]), and ankle (n=27 [15.8%]).

Conclusion: This study demonstrates that a web-based and competency-focused intervention can improve pMSK radiograph interpretation among R-PGTs and identifies cases prone to diagnostic error. These findings align with prior work showing the value of deliberate practice in radiology education.

Background: Deep learning image reconstruction (DLIR) has gained recognition as a promising technique to improve image quality in low-dose CT imaging. However, its performance in dual-energy CT portal venography (DE-CTPV), particularly under reduced contrast medium volume and radiation dose (dual-low dose) conditions, remains underexplored.

Objective: This study aims to compare the performance of DLIR and adaptive statistical iterative reconstruction (ASIR-V) in DE-CTPV, with a focus on image quality across multiple vascular segments of the portal venous (PV) system under dual-low dose protocols.

Methods: Patients undergoing DE-CTPV were reconstructed using DLIR medium (DLIR-M) and high strength (DLIR-H) and ASIR-V (50%). Image quality was assessed both subjectively and objectively in the main portal vein (MPV), left and right portal veins (LPV, RPV), splenic vein (SV), and superior mesenteric vein (SMV). Objective metrics, including image noise, contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR), were calculated. Additionally, radiation dose parameters (CTDI_vol, DLP, ED) and contrast medium volume were compared with data from previous studies.

Results: In this study, the mean CTDI_vol, DLP, and ED were 9.79 ± 2.13 mGy, 326.26 ± 84.58 mGy·cm, and 4.89 ± 1.27 mSv, respectively. The mean contrast medium volume was 79.5 ± 11.4 mL. DLIR-H significantly enhanced image quality across all vascular segments, achieving substantial reductions in image noise and notable increases in CNR and SNR (P < 0.05). It also received the highest subjective ratings for overall image quality, image noise, vascular edge sharpness, and diagnostic confidence compared to ASIR-V 50%. The use of 55 keV virtual monoenergetic imaging (VMI) further enhanced iodine contrast effectiveness, while DLIR effectively reduced noise, ensuring clearer and more consistent vascular delineation across all assessed vascular segments.

Conclusion: DLIR substantially improves image quality in DE-CTPV compared with ASIR-V 50%, even when utilizing dual-low dose protocol. By providing consistent, high-quality imaging across multiple portal venous segments, DLIR may offers a safer and more reliable approach for preoperative evaluation and postoperative monitoring in liver transplantation.