Radiology-Artificial Intelligence最新文献

Purpose To develop and evaluate an automated system for extracting structured clinical information from unstructured radiology and pathology reports using open-weight language models (LMs) and retrieval-augmented generation (RAG) and to assess the effects of model configuration variables on extraction performance. Materials and Methods This retrospective study used two datasets: 7294 radiology reports annotated for Brain Tumor Reporting and Data System (BT-RADS) scores and 2154 pathology reports annotated for IDH mutation status (January 2017-July 2021). An automated pipeline was developed to benchmark the performance of various LMs and RAG configurations for accuracy of structured data extraction from reports. The effect of model size, quantization, prompting strategies, output formatting, and inference parameters on model accuracy was systematically evaluated. Results The best-performing models achieved up to 98% accuracy in extracting BT-RADS scores from radiology reports and greater than 90% accuracy for extraction of IDH mutation status from pathology reports. The best model was medical fine-tuned Llama 3. Larger, newer, and domain fine-tuned models consistently outperformed older and smaller models (mean accuracy, 86% vs 75%; P < .001). Model quantization had minimal effect on performance. Few-shot prompting significantly improved accuracy (mean [±SD] increase, 32% ± 32; P = .02). RAG improved performance for complex pathology reports by a mean of 48% ± 11 (P = .001) but not for shorter radiology reports (-8% ± 31; P = .39). Conclusion This study demonstrates the potential of open LMs in automated extraction of structured clinical data from unstructured clinical reports with local privacy-preserving application. Careful model selection, prompt engineering, and semiautomated optimization using annotated data are critical for optimal performance. Keywords: Large Language Models, Retrieval-Augmented Generation, Radiology, Pathology, Health Care Reports Supplemental material is available for this article. © RSNA, 2025 See also commentary by Tejani and Rauschecker in this issue.

Purpose To develop an unsupervised deep learning framework for generalizable blood-brain barrier leakage detection using dynamic contrast-enhanced MRI, without requiring pharmacokinetic models and arterial input function estimation. Materials and Methods This retrospective study included data from patients who underwent dynamic contrast-enhanced MRI between April 2010 and December 2020. An autoencoder-based anomaly detection approach identified one-dimensional voxel-wise time-series abnormal signals through reconstruction residuals, separating them into residual leakage signals (RLSs) and residual vascular signals. The RLS maps were evaluated and compared with the volume transfer constant (K^trans) using the structural similarity index and correlation coefficient. Generalizability was tested on subsampled data, and isocitrate dehydrogenase (IDH) status classification performance was assessed using area under the receiver operating characteristic curve (AUC). Results A total of 274 patients (mean age, 54.4 years ± 14.6 [SD]; 164 male) were included in the study. RLS showed high structural similarity (structural similarity index, 0.91 ± 0.02) and correlation (r = 0.56; P < .001) with K^trans. On subsampled data, RLS maps showed better correlation with RLS values from the original data (0.89 vs 0.72; P < .001), higher peak signal-to-noise ratio (33.09 dB vs 28.94 dB; P < .001), and higher structural similarity index (0.92 vs 0.87; P < .001) compared with K^trans maps. RLS maps also outperformed K^trans maps in predicting IDH mutation status (AUC, 0.87 [95% CI: 0.83, 0.91] vs 0.81 [95% CI: 0.76, 0.85]; P = .02). Conclusion The unsupervised framework effectively detected blood-brain barrier leakage without pharmacokinetic models and arterial input function. Keywords: Dynamic Contrast-enhanced MRI, Unsupervised Learning, Feature Detection, Blood-Brain Barrier Leakage Detection Supplemental material is available for this article. © RSNA, 2025 See also commentary by Júdice de Mattos Farina and Kuriki in this issue.

Purpose To evaluate cancer detection and marker placement accuracy of two artificial intelligence (AI) models developed for interpretation of screening mammograms. Materials and Methods This retrospective study included data from 129 434 screening examinations (all female patients; mean age, 59.2 years ± 5.8 [SD]) performed between January 2008 and December 2018 in BreastScreen Norway. Model A was commercially available and model B was an in-house model. Area under the receiver operating characteristic curve (AUC) with 95% CIs were calculated. The study defined 3.2% and 11.1% of the examinations with the highest AI scores as positive, threshold 1 and 2, respectively. A radiologic review assessed location of AI markings and classified interval cancers as true or false negative. Results The AUC value was 0.93 (95% CI: 0.92, 0.94) for model A and B when including screen-detected and interval cancers. Model A identified 82.5% (611 of 741) of the screen-detected cancers at threshold 1 and 92.4% (685 of 741) at threshold 2. Model B identified 81.8% (606 of 741) at threshold 1 and 93.7% (694 of 741) at threshold 2. The AI markings were correctly localized for all screen-detected cancers identified by both models and 82% (56 of 68) of the interval cancers for model A and 79% (54 of 68) for model B. At the review, 21.6% (45 of 208) of the interval cancers were identified at the preceding screening by either or both models, correctly localized and classified as false negative (n = 17) or with minimal signs of malignancy (n = 28). Conclusion Both AI models showed promising performance for cancer detection on screening mammograms. The AI markings corresponded well to the true cancer locations. Keywords: Breast, Mammography, Screening, Computed-aided Diagnosis Supplemental material is available for this article. © RSNA, 2025 See also commentary by Cadrin-Chênevert in this issue.

Purpose To develop an adaptive dual-task deep learning model (ThyNet-S) for detection and classification of thyroid lesions at US screening. Materials and Methods This retrospective study used a multicenter dataset comprising 35 008 thyroid US images of 23 294 individual examinations (mean age, 40.4 years ± 13.1 [SD]; 17 587 female) from seven medical centers from January 2009 to December 2021. Of these, 29 004 images were used for model development and 6004 images for validation. The model determined cancer risk for each image and automatically triaged images with normal thyroid and benign nodules by dynamically integrating lesion detection through pixel-level feature analysis and lesion classification through deep semantic features analysis. Diagnostic performance of screening assisted by the model (ThyNet-S triaged screening) and traditional screening (radiologists alone) was assessed by comparing sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve using the McNemar test and DeLong test. The influence of ThyNet-S on radiologist workload and clinical decision-making was also assessed. Results ThyNet-S-assisted triaged screening achieved a higher area under the receiver operating characteristic curve than original screening with six senior and six junior radiologists (0.93 vs 0.91 and 0.92 vs 0.88, respectively; all P < .001). The model improved sensitivity for junior radiologists (88.2% vs 86.8%; P < .001). Notably, the model reduced radiologists' workload by triaging 60.4% of cases as not potentially malignant, which did not require further interpretation. The model simultaneously decreased the unnecessary fine needle aspiration rate from 38.7% to 14.9% and 11.5% when used independently or in combination with the Thyroid Imaging Reporting and Data System, respectively. Conclusion ThyNet-S improved the efficiency of thyroid cancer screening and optimized clinical decision-making. Keywords: Artificial Intelligence, Adaptive, Dual Task, Thyroid Cancer, Screening, Ultrasound Supplemental material is available for this article. © RSNA, 2025.

Purpose To develop a longitudinally aware segmentation network (LAS-Net) that can quantify serial PET/CT images for pediatric patients with Hodgkin lymphoma. Materials and Methods This retrospective study included baseline (PET1) and interim (PET2) PET/CT images from 297 pediatric patients enrolled in two Children's Oncology Group clinical trials (AHOD1331 and AHOD0831). The internal dataset included 200 patients (enrolled between March 2015 and August 2019; median age, 15.4 years [range, 5.6-22.0 years]; 107 male), and the external testing dataset included 97 patients (enrolled between December 2009 and January 2012; median age, 15.8 years [range, 5.2-21.4 years]; 59 male). LAS-Net incorporates longitudinal cross-attention, allowing relevant features from PET1 to inform the analysis of PET2. The model's lesion segmentation performance on PET1 images was evaluated using Dice coefficients, and lesion detection performance on PET2 images was evaluated using F1 scores. In addition, quantitative PET metrics, including metabolic tumor volume (MTV) and total lesion glycolysis (TLG) in PET1, as well as qPET and percentage difference between baseline and interim maximum standardized uptake value (∆SUV_max) in PET2, were extracted and compared against physician-derived measurements. Agreement between model and physician-derived measurements was quantified using Spearman correlation, and bootstrap resampling was used for statistical analysis. Results LAS-Net detected residual lymphoma on PET2 scans with an F1 score of 0.61 (precision/recall: 0.62/0.60), outperforming all comparator methods (P < .01). For baseline segmentation, LAS-Net achieved a mean Dice score of 0.77. In PET quantification, LAS-Net's measurements of qPET, ∆SUV_max, MTV, and TLG were strongly correlated with physician measurements, with Spearman ρ values of 0.78, 0.80, 0.93, and 0.96, respectively. The quantification performance remained high, with a slight decrease, in an external testing cohort. Conclusion LAS-Net demonstrated significant improvements in quantifying PET metrics across serial scans in pediatric patients with Hodgkin lymphoma, highlighting the value of longitudinal awareness in evaluating multi-time-point imaging datasets. Keywords: Pediatrics, PET/CT, Lymphoma, Segmentation, Quantification, Supervised Learning, Convolutional Neural Network (CNN), Quantitative PET, Longitudinal Analysis, Deep Learning, Image Segmentation Supplemental material is available for this article. Clinical trial registration no. NCT02166463 and NCT01026220 © RSNA, 2025 See also commentary by Khosravi and Gichoya in this issue.

Retrieval-augmented generation (RAG) is a strategy to improve the performance of large language models (LLMs) by providing an LLM with an updated corpus of knowledge that can be used for answer generation in real time. RAG may improve LLM performance and clinical applicability in radiology by providing citable, up-to-date information without requiring model fine-tuning. In this retrospective study, a radiology-specific RAG system was developed using a vector database of 3689 RadioGraphics articles published from January 1999 to December 2023. Performance of five LLMs with (RAG-Systems) and without RAG on a 192-question radiology examination was compared. RAG significantly improved examination scores for GPT-4 (OpenAI; 81.2% vs 75.5%, P = .04) and Command R+ (Cohere; 70.3% vs 62.0%, P = .02), but not for Claude Opus (Anthropic), Mixtral (Mistral AI), or Gemini 1.5 Pro (Google DeepMind). RAG-Systems performed significantly better than pure LLMs on a 24-question subset directly sourced from RadioGraphics (85% vs 76%, P = .03). RAG-Systems retrieved 21 of 24 (87.5%, P < .001) relevant RadioGraphics references cited in the examination's answer explanations and successfully cited them in 18 of 21 (85.7%, P < .001) outputs. The results suggest that RAG is a promising approach to enhance LLM capabilities for radiology knowledge tasks, providing transparent, domain-specific information retrieval. Keywords: Computer Applications-General (Informatics), Technology Assessment Supplemental material is available for this article. © RSNA, 2025 See also commentary by Mansuri and Gichoya in this issue.