Korean Journal of Radiology最新文献

Objective: To evaluate the accuracy of large language models (LLMs) in extracting Coronary Artery Disease-Reporting and Data System (CAD-RADS) 2.0 components from coronary CT angiography (CCTA) reports, and assess the impact of prompting strategies.

Materials and methods: In this multi-institutional study, we collected 319 synthetic, semi-structured CCTA reports from six institutions to protect patient privacy while maintaining clinical relevance. The dataset included 150 reports from a primary institution (100 for instruction development and 50 for internal testing) and 169 reports from five external institutions for external testing. Board-certified radiologists established reference standards following the CAD-RADS 2.0 guidelines for all three components: stenosis severity, plaque burden, and modifiers. Six LLMs (GPT-4, GPT-4o, Claude-3.5-Sonnet, o1-mini, Gemini-1.5-Pro, and DeepSeek-R1-Distill-Qwen-14B) were evaluated using an optimized instruction with prompting strategies, including zero-shot or few-shot with or without chain-of-thought (CoT) prompting. The accuracy was assessed and compared using McNemar's test.

Results: LLMs demonstrated robust accuracy across all CAD-RADS 2.0 components. Peak stenosis severity accuracies reached 0.980 (48/49, Claude-3.5-Sonnet and o1-mini) in internal testing and 0.946 (158/167, GPT-4o and o1-mini) in external testing. Plaque burden extraction showed exceptional accuracy, with multiple models achieving perfect accuracy (43/43) in internal testing and 0.993 (137/138, GPT-4o, and o1-mini) in external testing. Modifier detection demonstrated consistently high accuracy (≥0.990) across most models. One open-source model, DeepSeek-R1-Distill-Qwen-14B, showed a relatively low accuracy for stenosis severity: 0.898 (44/49, internal) and 0.820 (137/167, external). CoT prompting significantly enhanced the accuracy of several models, with GPT-4 showing the most substantial improvements: stenosis severity accuracy increased by 0.192 (P < 0.001) and plaque burden accuracy by 0.152 (P < 0.001) in external testing.

Conclusion: LLMs demonstrated high accuracy in automated extraction of CAD-RADS 2.0 components from semi-structured CCTA reports, particularly when used with CoT prompting.

Objective: To evaluate the accuracy of multimodal large language models (LLMs) in detecting cases requiring immediate radiology reporting in pediatric radiology.

Materials and methods: Seventy-one publicly available, paraphrased pediatric clinical vignettes with images-sourced from the New England Journal of Medicine, The Lancet, Archives of Pediatrics & Adolescent Medicine, and Radiology-were assessed by seven vision-capable LLMs (temperature levels 0 and 1; t0 and t1) and four human readers (an expert pediatric radiologist, a trainee radiologist, an expert pediatrician, and a trainee pediatrician). Cases were classified as requiring immediate reporting (n = 33) if they corresponded to Korean Triage and Acuity Scale (KTAS) levels 1-2 (n = 24) or met the criteria for a critical value report (CVR) (n = 11). The most accurate LLM was compared with each human reader, with significance set at P < 0.013.

Results: LLMs demonstrated 60.6%-83.1% accuracy in detecting cases requiring immediate radiology reporting: 57.7%-81.7% and 53.5%-87.3% for KTAS levels 1-2 and CVR cases, respectively. Gemini-Flash with t1 showed the highest accuracy among the LLMs: 83.1% (95% confidence interval [CI]: 74.6%-91.5%), 81.7% (95% CI: 71.8%-90.1%), and 87.3% (95% CI: 78.9%-94.4%) for identifying cases requiring immediate reporting, KTAS level 1-2 cases, and CVR cases, respectively, despite its low sensitivity for CVR detection (3/11, 27.3%). Human readers demonstrated 62.0%-84.5% accuracy for immediate radiology reporting, 73.2%-84.5% for KTAS levels 1-2, and 39.4%-94.4% for CVR cases. The accuracy of Gemini-Flash t1 in identifying cases requiring immediate radiology reporting was comparable to that of the most accurate human reader (vs. expert pediatrician: 84.5% [95% CI: 76.1%-93.0%]; P < 0.99).

Conclusion: Multimodal LLMs may achieve overall accuracy comparable to or exceeding that of human readers in identifying cases requiring immediate radiology reporting, supporting their potential use for pediatric radiology worklist prioritization. However, the models' sensitivity in detecting such cases was not reliable.

Ovarian cancer (OC) remains one of the leading causes of gynecologic cancer-related mortality, with most patients presenting with disseminated disease, particularly within the peritoneal cavity. Standard treatment includes cytoreductive surgery, platinum-based chemotherapy, and targeted maintenance approaches depending on the patient's and tumor's genetic profile. Despite treatment advancements, approximately 25% of high-grade serous OC cases relapse within a year despite optimal primary treatment with complete tumor clearance at cytoreduction. Advances in contrast-enhanced CT (CE-CT) and MRI have revolutionized the evaluation and treatment planning of advanced OC. CT remains the gold standard for staging and assessing tumor extent, effectively identifying peritoneal, lymphatic, and distant metastases. However, it is less effective in detecting small-volume peritoneal dissemination. MRI, with superior soft-tissue contrast, complements CT by providing a detailed assessment of peritoneal disease, characterizing sonographically indeterminate adnexal masses. Diffusion-weighted imaging and gadolinium-enhanced MRI have improved the diagnostic sensitivity for peritoneal disease but are unable to predict treatment response, recurrence risk, and prognosis. Radiomics, which extracts quantitative tumor features from imaging data, holds promise for personalizing treatment and identifying patients at risk for early recurrence despite optimal therapy. The integration of CT, MRI, and radiomics could enhance surgical planning and improve long-term survival outcomes in patients with advanced OC.

Objective: To develop a deep learning model for estimating newborn gestational age (GA) based on the shape of the lumbar vertebral bodies on cross-table lateral radiographs obtained on the first day after birth.

Materials and methods: This retrospective study included 423 cross-table lateral radiographs of 423 newborns (242 boys and 181 girls) taken within 24 hours after birth at two hospitals. Of these, 256 radiographs (157 boys and 99 girls) obtained from one institution were used for model development, and 167 radiographs (85 boys and 82 girls) from the other institution were used for model external testing. Clinical data, including medical history of underlying disorders, GA determined by ultrasound parameters, birth date, birth weight, sex, examination date, and reason for requesting radiographs, were obtained. The radiographs underwent manual labeling of the five lumbar vertebral bodies, followed by preprocessing steps such as normalization, resizing, denoising, cropping, and augmentation via horizontal flipping and rotation. Subsequently, we trained a deep learning model using a DeepLabv3+ network with a ResNet50 backbone for lumbar segmentation and used a customized AgeClassifier model with two parallel ResNet18 backbones for GA estimation. Model performance was evaluated using an external test dataset after image cropping.

Results: Neither GA nor birth weight differed significantly between boys and girls. In the segmentation model, the mean dice similarity coefficient ± standard deviation (SD) was 0.801 ± 0.031. For GA estimation, the mean absolute error ± SD was 5.2 ± 0.5 days. The Bland-Altman bias (AI-estimated GA - ground truth GA) and 95% limits of agreement were -0.4 days and -13.0 to 12.3 days, respectively.

Conclusion: Our deep learning model showed promising performance in lumbar vertebral body segmentation and GA estimation using plain radiographs, suggesting its potential utility as a supportive tool for neonatal maturity assessment in clinical practice.

In 2022, nearly 2.3 million new cases of breast cancer were reported globally, with less than half of these cases originating from Asia. Despite the relatively low incidence of breast cancer in most parts of Asia, the mortality-to-incidence ratio remains high. Low-income countries lack resources for breast cancer screening, whereas high-income countries fail to fully benefit from national breast screening programs because of the underutilization of preventive healthcare services. There is a notable difference in the age distribution of breast cancer cases between Asian and Western populations, with the prevalence peaking approximately a decade earlier in Asian women and most commonly affecting those aged 40-50 years. Existing literature on breast cancer trends, screening guidelines, and clinical practices in Asian countries, particularly regarding regional variations and healthcare system differences, is relatively sparse. Gaining a deeper understanding of how different Asian countries are implementing breast cancer screening in response to the rising incidence of the disease can help identify tailored strategies for early detection, ultimately contributing to a reduction in breast cancer-related mortality. This review explored the current breast cancer landscape, including breast cancer screening guidelines and outcomes of screening examinations in Asia, highlighting key challenges and future directions.

Objective: To evaluate the association between hypoxic-ischemic injury (HII) of the brain and glymphatic function using MRI-derived parameters in neonates.

Materials and methods: This retrospective, single-institution study collected brain MRI scans of 127 neonates between July 2020 and July 2022. The volume and fraction of the basal ganglia perivascular space (BG-PVS) were automatically extracted using three-dimensional T2-weighted image processing. Diffusion-tensor imaging (DTI) along the PVS (DTI-ALPS) index values were derived from the DTI maps. BG-PVS and DTI-ALPS parameters were compared between neonates with and without HII. The correlations between MRI-derived glymphatic parameters and corrected gestational age (CGA), as well as between BG-PVS measurements and the DTI-ALPS index, were analyzed using Spearman coefficients. Multivariable logistic regression adjusted for age, sex, birth weight, and mode of delivery was performed to examine the associations between each glymphatic parameter and HII.

Results: This study included 97 neonates without HII (median gestational age [GA]: 252 days) and 30 with HII (median GA: 252 days). Neonates with HII had smaller BG-PVS volumes (19 mm³ vs. 33 mm³, P = 0.001) and fractions (0.29% vs. 0.54%, P = 0.003) compared to neonates without HII. The DTI-ALPS index values did not differ significantly between neonates with and without HII (P = 0.54). CGA correlated negatively with BG-PVS measurements (ρ = -0.21 to -0.26, all P < 0.05) and positively with DTI-ALPS index values (ρ = 0.22, P = 0.014). BG-PVS measurements and DTI-ALPS index values were not significantly correlated (ρ = -0.28 to -0.08, all P > 0.05). Multivariable logistic regression revealed a negative association between BG-PVS volume (odds ratio [OR]: 0.96 per mm³ increase, 95% confidence interval [CI]: 0.93-0.99) and fraction (OR: 0.15 per % increase, 95% CI: 0.03-0.79) with HII, while DTI-ALPS index values were not significantly associated with HII (OR: 0.10, 95% CI: 0.00-25.41).

Conclusion: Neonates with HII demonstrated smaller BG-PVS volume and fraction compared with those without HII, indicating potential alterations in glymphatic function among affected newborns.

Objective: To evaluate the effect of deep learning (DL)-based artificial intelligence (AI) software on the diagnostic performance of radiologists with different experience levels in detecting nigrosome 1 (N1) abnormalities on susceptibility map-weighted imaging (SMwI).

Materials and methods: This retrospective diagnostic case-control study analyzed 139 SMwI scans of 59 patients with Parkinson's disease (PD) and 80 healthy participants. Participants were imaged using 3T MRI, and AI-generated assessments for N1 abnormalities were obtained using an AI model (version 1.0.1.0; Heuron Corporation, Seoul, Korea), which utilized YOLOX-based object detection and SparseInst segmentation models. Four radiologists (two experienced neuroradiologists and two less experienced residents) evaluated N1 abnormalities with and without AI in a crossover study design. Diagnostic performance metrics, inter-reader agreements, and reader responses to AI-generated assessments were evaluated.

Results: Use of AI significantly improved diagnostic performance compared with interpretation without it across three readers, with significant increases in specificity (0.86 vs. 0.94, P = 0.004; 0.91 vs. 0.97, P = 0.024; and 0.90 vs. 0.97, P = 0.012). Inter-reader agreement also improved with AI, as Fleiss's kappa increased from 0.73 (95% confidence interval [CI]: 0.61-0.84) to 0.87 (95% CI: 0.76-0.99). The net reclassification index (NRI) demonstrated significant improvement in three of the four readers. When grouped by experience level, less experienced readers showed greater improvement (NRI = 12.8%, 95% CI: 0.067-0.190) than experienced readers (NRI = 0.8%, 95% CI: -0.037-0.051). In the less experienced group, reader-AI disagreement was significantly higher in the PD group than in the normal group (8.1% vs. 3.8%, P = 0.029).

Conclusion: DL-based AI enhances the diagnostic performance in detecting N1 abnormalities on SMwI, particularly benefiting less experienced radiologists. These findings underscore the potential for improving diagnostic workflows for PD.

Objective: To evaluate the impact of deep learning-based image conversion on the accuracy of automated coronary artery calcium quantification using thin-slice, sharp-kernel, non-gated, low-dose chest computed tomography (LDCT) images collected from multiple institutions.

Materials and methods: A total of 225 pairs of LDCT and calcium scoring CT (CSCT) images scanned at 120 kVp and acquired from the same patient within a 6-month interval were retrospectively collected from four institutions. Image conversion was performed for LDCT images using proprietary software programs to simulate conventional CSCT. This process included 1) deep learning-based kernel conversion of low-dose, high-frequency, sharp kernels to simulate standard-dose, low-frequency kernels, and 2) thickness conversion using the raysum method to convert 1-mm or 1.25-mm thickness images to 3-mm thickness. Automated Agaston scoring was conducted on the LDCT scans before (LDCT-Org_auto) and after the image conversion (LDCT-CONV_auto). Manual scoring was performed on the CSCT images (CSCT_manual) and used as a reference standard. The accuracy of automated Agaston scores and risk severity categorization based on the automated scoring on LDCT scans was analyzed compared to the reference standard, using the Bland-Altman analysis, concordance correlation coefficient (CCC), and weighted kappa (κ) statistic.

Results: LDCT-CONV_auto demonstrated a reduced bias for Agaston score, compared with CSCT_manual, than LDCT-Org_auto did (-3.45 vs. 206.7). LDCT-CONV_auto showed a higher CCC than LDCT-Org_auto did (0.881 [95% confidence interval {CI}, 0.750-0.960] vs. 0.269 [95% CI, 0.129-0.430]). In terms of risk category assignment, LDCT-Org_auto exhibited poor agreement with CSCT_manual (weighted κ = 0.115 [95% CI, 0.082-0.154]), whereas LDCT-CONV_auto achieved good agreement (weighted κ = 0.792 [95% CI, 0.731-0.847]).

Conclusion: Deep learning-based conversion of LDCT images originally obtained with thin slices and a sharp kernel can enhance the accuracy of automated coronary artery calcium score measurement using the images.