Radiology最新文献

History: A 38-year-old previously healthy male patient presented with left-sided facial pain over the prior 5 weeks. He first noticed the pain while washing and applying pressure to his face. The pain was described as shock-like, sharp and shooting, and radiating along the left cheek and temple. It began as 1-2-second episodes occurring two to three times per day, sometimes spontaneously, progressing in severity and frequency over time. Mild progressive left facial weakness also developed a few weeks after initial symptoms. Physical examination demonstrated reproducible pain in the distribution of the maxillary division of the trigeminal nerve (V2), with normal motor and sensory function. A recent routine dental examination demonstrated healthy teeth and gums, and there was no history of dental procedures or trauma. The rest of the physical and neurologic examinations revealed no abnormalities. The patient was afebrile with normal vital signs. Findings of routine laboratory testing, including complete blood count, metabolic panel with electrolytes, kidney and liver function, and inflammatory markers such as C-reactive protein, were all within normal limits. Following the neurologic and otolaryngologic evaluations, imaging was recommended. The patient was also started on treatment with carbamazepine for trigeminal neuralgia, with modest improvement of symptoms. He initially underwent MRI of the temporal bones at an outside hospital. After subsequent referral to our hospital, follow-up concomitant MRI and CT (Figs 1-4) were performed approximately 3 months after the initial imaging.

Integrating large language models (LLMs) into health care holds substantial potential to enhance clinical workflows and care delivery. However, LLMs also pose serious risks if integration is not thoughtfully executed, with complex challenges spanning accuracy, accessibility, privacy, and regulation. Proprietary commercial LLMs (eg, GPT-4 [OpenAI], Claude 3 Sonnet and Claude 3 Opus [Anthropic], Gemini [Google]) have received much attention from researchers in the medical domain, including radiology. Interestingly, open-source LLMs (eg, Llama 3 and LLaVA-Med) have received comparatively little attention. Yet, open-source LLMs hold several key advantages over proprietary LLMs for medical institutions, hospitals, and individual researchers. The wider adoption of open-source LLMs has been slower, perhaps in part due to the lack of familiarity, accessible computational infrastructure, and community-built tools to streamline their local implementation and customize them for specific use cases. Thus, this article provides a tutorial for the implementation of open-source LLMs in radiology, including examples of commonly used tools for text generation and techniques for troubleshooting issues with prompt engineering, retrieval-augmented generation, and fine-tuning. Implementation-ready code for each tool is provided at https://github.com/UM2ii/Open-Source-LLM-Tools-for-Radiology. In addition, this article compares the benefits and drawbacks of open-source and proprietary LLMs, discusses the differentiating characteristics of popular open-source LLMs, and highlights recent advancements that may affect their adoption.

Background The ninth edition of the TNM classification for lung cancer revised the N2 categorization, improving patient stratification, but prognostic heterogeneity remains for the N1 category. Purpose To define the optimal size cutoff for a bulky lymph node (LN) on CT scans and to evaluate the prognostic value of bulky LN in the clinical N staging of lung cancer. Materials and Methods This retrospective study analyzed patients who underwent lobectomy or pneumonectomy for lung cancer between January 2013 and December 2021, divided into development (2016-2021) and validation (2013-2015) cohorts. The optimal threshold for a bulky LN was defined based on the short-axis diameter of the largest clinically positive LN at CT. Prognostic differences according to presence of bulky LN in cN1 category for overall survival (OS) were evaluated using multivariable Cox analysis. Survival discrimination was assessed using the Harrell concordance index (C-index). Results A total of 3426 patients (mean age, 64.0 years ± 9.3 [SD]; 1837 male) and 1327 patients (mean age, 63.0 years ± 9.7; 813 male) were included in the development and validation cohorts, respectively. The cutoff size for a bulky LN was established at 15 mm, and the presence of bulky LN was an independent risk factor for OS (hazard ratio [HR], 1.54; 95% CI: 1.10, 2.16; P = .01). In the development and validation cohorts, the cN1-bulky group had higher mortality risk than the cN1-nonbulky group (HR, 2.82 [95% CI: 1.73, 4.58; P < .001]; 2.29 [95% CI: 1.34, 3.92; P = .002], respectively). The bulky LN descriptor improved prognostic discrimination within the cN1 category compared with the current staging (C-index from 0.50 to 0.60 and to 0.58 in the development and validation cohorts [P < .001, P = .006], respectively]). Conclusion Defining bulky LN with a size cutoff of 15 mm was an effective descriptor in the clinical staging of N1 lung cancer. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by Horst in this issue.

Background Large-scale secondary use of clinical databases requires automated tools for retrospective extraction of structured content from free-text radiology reports. Purpose To share data and insights on the application of privacy-preserving open-weights large language models (LLMs) for reporting content extraction with comparison to standard rule-based systems and the closed-weights LLMs from OpenAI. Materials and Methods In this retrospective exploratory study conducted between May 2024 and September 2024, zero-shot prompting of 17 open-weights LLMs was preformed. These LLMs with model weights released under open licenses were compared with rule-based annotation and with OpenAI's GPT-4o, GPT-4o-mini, GPT-4-turbo, and GPT-3.5-turbo on a manually annotated public English chest radiography dataset (Indiana University, 3927 patients and reports). An annotated nonpublic German chest radiography dataset (18 500 reports, 16 844 patients [10 340 male; mean age, 62.6 years ± 21.5 {SD}]) was used to compare local fine-tuning of all open-weights LLMs via low-rank adaptation and 4-bit quantization to bidirectional encoder representations from transformers (BERT) with different subsets of reports (from 10 to 14 580). Nonoverlapping 95% CIs of macro-averaged F1 scores were defined as relevant differences. Results For the English reports, the highest zero-shot macro-averaged F1 score was observed for GPT-4o (92.4% [95% CI: 87.9, 95.9]); GPT-4o outperformed the rule-based CheXpert [Stanford University] (73.1% [95% CI: 65.1, 79.7]) but was comparable in performance to several open-weights LLMs (top three: Mistral-Large [Mistral AI], 92.6% [95% CI: 88.2, 96.0]; Llama-3.1-70b [Meta AI], 92.2% [95% CI: 87.1, 95.8]; and Llama-3.1-405b [Meta AI]: 90.3% [95% CI: 84.6, 94.5]). For the German reports, Mistral-Large (91.6% [95% CI: 90.5, 92.7]) had the highest zero-shot macro-averaged F1 score among the six other open-weights LLMs and outperformed the rule-based annotation (74.8% [95% CI: 73.3, 76.1]). Using 1000 reports for fine-tuning, all LLMs (top three: Mistral-Large, 94.3% [95% CI: 93.5, 95.2]; OpenBioLLM-70b [Saama]: 93.9% [95% CI: 92.9, 94.8]; and Mixtral-8×22b [Mistral AI]: 93.8% [95% CI: 92.8, 94.7]) achieved significantly higher macro-averaged F1 score than did BERT (86.7% [95% CI: 85.0, 88.3]); however, the differences were not relevant when 2000 or more reports were used for fine-tuning. Conclusion LLMs have the potential to outperform rule-based systems for zero-shot "out-of-the-box" structuring of report databases, with privacy-ensuring open-weights LLMs being competitive with closed-weights GPT-4o. Additionally, the open-weights LLM outperformed BERT when moderate numbers of reports were used for fine-tuning. Published under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Gee and Yao in this issue.

Background Studies have explored the application of multimodal large language models (LLMs) in radiologic differential diagnosis. Yet, how different multimodal input combinations affect diagnostic performance is not well understood. Purpose To evaluate the impact of varying multimodal input elements on the accuracy of OpenAI's GPT-4 with vision (GPT-4V)-based brain MRI differential diagnosis. Materials and Methods Sixty brain MRI cases with a challenging yet verified diagnosis were selected. Seven prompt groups with variations of four input elements (image without modifiers [I], annotation [A], medical history [H], and image description [D]) were defined. For each MRI case and prompt group, three identical queries were performed using an LLM-based search engine (Perplexity AI, powered by GPT-4V). The accuracy of LLM-generated differential diagnoses was rated using a binary and a numeric scoring system and analyzed using a χ² test and a Kruskal-Wallis test. Results were corrected for false-discovery rate with use of the Benjamini-Hochberg procedure. Regression analyses were performed to determine the contribution of each input element to diagnostic performance. Results The prompt group containing I, A, H, and D as input exhibited the highest diagnostic accuracy (124 of 180 responses [69%]). Significant differences were observed between prompt groups that contained D among their inputs and those that did not. Unannotated (I) (four of 180 responses [2.2%]) or annotated radiologic images alone (I and A) (two of 180 responses [1.1%]) yielded very low diagnostic accuracy. Regression analyses confirmed a large positive effect of D on diagnostic accuracy (odds ratio [OR], 68.03; P < .001), as well as a moderate positive effect of H (OR, 4.18; P < .001). Conclusion The textual description of radiologic image findings was identified as the strongest contributor to the performance of GPT-4V in brain MRI differential diagnosis, followed by the medical history; unannotated or annotated images alone yielded very low diagnostic performance. © RSNA, 2025 Supplemental material is available for this article.