Purpose: This observational study aims to provide a detailed clinical and imaging characterization/workup of acute intracerebral hemorrhage (ICH) due to either an underlying metastasis (mICH) or brain tumor (tICH) lesion. Methods: We conducted a retrospective, single-center study, evaluating patients presenting with occult ICH on initial CT imaging, classified as tICH or mICH on follow-up MRI imaging according to the H-Atomic classification. Demographic, clinical and radiological data were reviewed. Results: We included 116 patients (tICH: 20/116, 17.24%; mICH: 96/116, 82.76%). The most common malignancies causing ICH were lung cancer (27.59%), malignant melanoma (18.10%) and glioblastoma (10.34%). The three most common stroke-like symptoms observed were focal deficit (62/116, 53.45%), dizziness (42/116, 36.21%) and cognitive impairment (27/116, 23.28%). Highest mICH prevalence was seen in the occipital lobe (mICH: 28.13%, tICH: 0.00%; p = 0.004) with tICH more in the corpus callosum (tICH: 10.00%, mICH: 0.00%; p = 0.029). Anticoagulation therapy was only frequent in mICH patients (tICH: 0.00%, mICH: 5.21%; p = 0.586). Hemorrhage (tICH: 12682 mm³, mICH: 5708 mm³, p = 0.020) and edema volumes (tICH: 49389 mm³, mICH: 20972 mm³, p = 0.035) were significantly larger within tICH patients. Conclusion: More than half of the patients with neoplastic ICH exhibited stroke-like symptoms. Lung cancer was most common in mICH, glioblastoma in tICH. While clinical presentations were similar, significant differences in tumor location and treatments were discernible.

Radiology reports can change during workflows, especially when residents draft preliminary versions that attending physicians finalize. We explored how large language models (LLMs) and embedding techniques can categorize these changes into textual, semantic, or clinically actionable types. We evaluated 400 adult CT reports drafted by residents against finalized versions by attending physicians. Changes were rated on a five-point scale from no changes to critical ones. We examined open-source LLMs alongside traditional metrics like normalized word differences, Levenshtein and Jaccard similarity, and text embedding similarity. Model performance was assessed using quadratic weighted Cohen's kappa (κ), (balanced) accuracy, F₁, precision, and recall. Inter-rater reliability among evaluators was excellent (κ = 0.990). Of the reports analyzed, 1.3 % contained critical changes. The tested methods showed significant performance differences (P < 0.001). The Qwen3-235B-A22B model using a zero-shot prompt, most closely aligned with human assessments of changes in clinical reports, achieving a κ of 0.822 (SD 0.031). The best conventional metric, word difference, had a κ of 0.732 (SD 0.048), the difference between the two showed statistical significance in unadjusted post-hoc tests (P = 0.038) but lost significance after adjusting for multiple testing (P = 0.064). Embedding models underperformed compared to LLMs and classical methods, showing statistical significance in most cases. Large language models like Qwen3-235B-A22B demonstrated moderate to strong alignment with expert evaluations of the clinical significance of changes in radiology reports. LLMs outperformed embedding methods and traditional string and word approaches, achieving statistical significance in most instances. This demonstrates their potential as tools to support peer review.

Purpose: This study evaluates seven open-source Large Language Models (LLMs) in summarizing radiology reports of acute ischemic stroke patients treated with mechanical thrombectomy and predicting angiography-based outcome measures relevant to post-thrombectomy reperfusion. Materials and methods: 2000 mechanical thrombectomy reports (findings and summarizing impression section as gold standard) were split into training set ( N = 1900) for model fine-tuning and test set ( N = 100). A two-step evaluation was performed: (1) Quantitative analyses of seven LLMs with metrics ROUGE-1, -2, -L, METEOR, BERTScore (F1) and BLEU comparing LLM-generated summaries against gold-standard impressions. (2) Qualitative manual evaluation of the four best-performing models by two radiologists, assessing correctness and completeness across key parameters: outcome-relevant scores, vessel information, occlusion side, number of passes, relevant additional information, hallucinations, and grammar quality. Statistical significance was assessed via a two-tailed, four-sample χ² test, followed by post hoc pairwise χ² comparisons. Results: BioMistral-7b scored highest across most quantitative metrics (ROUGE-1: 0.47, ROUGE-2: 0.30, ROUGE-L: 0.43, METEOR: 0.46, BERTScore (F1): 0.82). Manual evaluation revealed gemma-2-9b most frequently documented pass counts (56 out of 100 cases (56%); p < 0.02 vs. Llama-3.1-8b/mistral-7b-instruct), while mistral-7b-instruct described them most often correctly (29 out of 38 mentioned passes (76.32%); p < 0.02 vs. BioMistral-7b and p < 0.01 vs. gemma-2-9b). All four manually evaluated LLMs performed moderately well in predicting “Thrombolysis-In-Cerebral-Ischemia (TICI)” Score (correctness rate ranging from 66 to 71%; p = 0.89). Conclusion: All four manually evaluated LLMs effectively summarized thrombectomy reports and demonstrated moderate accuracy predicting TICI scores. Their integration into radiology workflows could enhance efficiency, warranting further clinical validation. Key Points: Question Specifically fine-tuned Large Language Models (LLMs) can improve radiology workflow by automatically summarizing thrombectomy reports and inferring angiographic classifications from textual descriptions. Findings Fine-tuned LLMs achieve similar performance in summarizing thrombectomy reports, with each model performing best in specific categories and showing moderate accuracy in correct “Thrombolysis-In-Cerebral-Ischemia (TICI)” Score prediction (66–71%). Clinical relevance Integrating fine-tuned LLMs into radiology workflows may accelerate decision-making and improve patient outcomes by automatically summarizing reports and assessing recanalization success, while future work should enhance contextual understanding, address ambiguous inputs, and limit hallucinations. Graphical

Background/Objectives: This study evaluates whether convolutional neural networks (CNNs) can be trained to determine the primary tumor origin from MRI images alone in patients with metastatic brain lesions. Methods: This retrospective, monocentric study involved the segmentation of 1175 brain lesions from MRI scans of 436 patients with histologically confirmed primary tumor origins. The four most common tumor types—lung adenocarcinoma, small cell lung cancer, breast cancer, and melanoma—were selected, and a class-balanced dataset was created through under-sampling. This resulted in 276 training datasets and 88 hold-out test datasets. Bayesian optimization was employed to determine the optimal CNN architecture, the most relevant imaging sequences, and whether the masking of images was necessary. We compared the performance of the CNN with that of two expert radiologists specializing in neuro-oncological imaging. Results: The best-performing CNN from the Bayesian optimization process used masked images across all available MRI sequences. It achieved Area-Under-the-Curve (AUC) values of 0.75 for melanoma, 0.65 for small cell lung cancer, 0.64 for breast cancer, and 0.57 for lung adenocarcinoma. Masked images likely improved performance by focusing the CNN on relevant regions and reducing noise from surrounding tissues. In comparison, Radiologist 1 achieved AUCs of 0.55, 0.52, 0.45, and 0.51, and Radiologist 2 achieved AUCs of 0.68, 0.55, 0.64, and 0.43 for the same tumor types, respectively. The CNN consistently showed higher accuracy, particularly for melanoma and breast cancer. Conclusions: Bayesian optimization enabled the creation of a CNN that outperformed expert radiologists in classifying the primary tumor origin of brain metastases from MRI.

Background/Objectives: This study compares the accuracy of responses from state-of-the-art large language models (LLMs) to patient questions before CT and MRI imaging. We aim to demonstrate the potential of LLMs in improving workflow efficiency, while also highlighting risks such as misinformation. Methods: There were 57 CT-related and 64 MRI-related patient questions displayed to ChatGPT-4o, Claude 3.5 Sonnet, Google Gemini, and Mistral Large 2. Each answer was evaluated by two board-certified radiologists and scored for accuracy/correctness/likelihood to mislead using a 5-point Likert scale. Statistics compared LLM performance across question categories. Results: ChatGPT-4o achieved the highest average scores for CT-related questions and tied with Claude 3.5 Sonnet for MRI-related questions, with higher scores across all models for MRI (ChatGPT-4o: CT [4.52 (± 0.46)], MRI: [4.79 (± 0.37)]; Google Gemini: CT [4.44 (± 0.58)]; MRI [4.68 (± 0.58)]; Claude 3.5 Sonnet: CT [4.40 (± 0.59)]; MRI [4.79 (± 0.37)]; Mistral Large 2: CT [4.25 (± 0.54)]; MRI [4.74 (± 0.47)]). At least one response per LLM was rated as inaccurate, with Google Gemini answering most often potentially misleading (in 5.26% for CT and 2.34% for MRI). Mistral Large 2 was outperformed by ChatGPT-4o for all CT-related questions (p < 0.001) and by ChatGPT-4o (p = 0.003), Google Gemini (p = 0.022), and Claude 3.5 Sonnet (p = 0.004) for all CT Contrast media information questions. Conclusions: Even though all LLMs performed well overall and showed great potential for patient education, each model occasionally displayed potentially misleading information, highlighting the clinical application risk.

Background: Radiology trainees require efficient, accurate, and accessible resources to master complex imaging techniques and identify findings that guide clinical decision-making. Large language models (LLMs) are emerging as promising tools for medical education and clinical workflows, offering the potential to enhance learning by providing instant feedback, aiding in diagnostic accuracy, and offering personalized learning experiences. However, systematic comparisons of LLMs for radiology education and clinical support remain limited, particularly regarding differences across subspecialties and resident experience levels. Objective: This study aimed to evaluate and compare the response quality of 2 state-of-the-art reasoning-based LLMs, namely DeepSeek-R1 and ChatGPT-o1, as clinical and radiology residency support tools, comparing performance across clinical and didactic dimensions, including text- and image-based responses. Methods: Overall, 27 radiology questions covering 9 radiological subspecialties were answered by both LLMs. Additionally, 6 image-based questions were presented only to ChatGPT-o1 due to its image processing capabilities. Responses were independently rated by 7 radiology residents (postgraduate years 2‐5) across 9 rating criteria grouped into 3 dimensions (factual accuracy, clinical practicality, and didactic value), using a 5-point Likert scale. Statistics compared LLMs, reader experience, and response types for text- as well as image-based for ChatGPT-o1 queries. Results: DeepSeek-R1 consistently outperformed ChatGPT-o1 across all rating dimensions, with highly significant differences across all criteria (mean ratings: DeepSeek-R1 4.51, SD 0.73 vs ChatGPT-o1 3.73, SD 0.98; P <.001). In an exploratory subspecialty-level analysis, DeepSeek-R1 descriptively outperformed ChatGPT-o1 across all subspecialties. For both LLMs accumulated, junior residents tended to rate slightly higher than seniors in 7 of 9 criteria, although differences were not statistically significant. However, for ChatGPT-o1, junior residents rated significantly higher in overall score across all criteria (juniors 3.81, SD 0.64 vs seniors 3.63, SD 0.65; P =.02). Image-based responses by ChatGPT-o1 scored significantly lower than text-based (mean 3.19, SD 1.42; P =.007), particularly in factual accuracy (mean 2.75, SD 1.45; P <.001) and clinical practicality (mean 3.11, SD 1.47; P =.03). Conclusions: Both DeepSeek-R1 and ChatGPT-o1 demonstrate promising potential on simulated radiology question sets designed for educational and clinical contexts, with DeepSeek-R1 outperforming ChatGPT-o1 across all evaluated criteria. These results emphasize the value of open-source models for educational use and provide early evidence that LLMs may support radiology resident training under controlled conditions; however, their real-world educational and clinical effects require further investigation. Future research should prospectively evaluate how LLMs can be integrated into radiology training, assess their impact alongside conventional teaching methods, and investigate multimodal capabilities to better reflect realistic clinical scenarios.