Several large language models showed high accuracy and quality for translating radiology reports across high-resource and low-resource languages; however, no one-size-fits-all model was identified.

Background A study was undertaken to assess the effectiveness of open-source large language models (LLMs) in extracting clinical data from unstructured mechanical thrombectomy reports in patients with ischemic stroke caused by a vessel occlusion. Methods We deployed local open-source LLMs to extract data points from free-text procedural reports in patients who underwent mechanical thrombectomy between September 2020 and June 2023 in our institution. The external dataset was obtained from a second university hospital and comprised consecutive cases treated between September 2023 and March 2024. Ground truth labeling was facilitated by a human-in-the-loop (HITL) approach, with time metrics recorded for both automated and manual data extractions. We tested three models—Mixtral, Qwen, and BioMistral—assessing their performance on precision, recall, and F1 score across 15 clinical categories such as National Institute of Health Stroke Scale (NIHSS) scores, occluded vessels, and medication details. Results The study included 1000 consecutive reports from our primary institution and 50 reports from a secondary institution. Mixtral showed the highest precision, achieving 0.99 for first series time extraction and 0.69 for occluded vessel identification within the internal dataset. In the external dataset, precision ranged from 1.00 for NIHSS scores to 0.70 for occluded vessels. Qwen showed moderate precision with a high of 0.85 for NIHSS scores and a low of 0.28 for occluded vessels. BioMistral had the broadest range of precision, from 0.81 for first series times to 0.14 for medication details. The HITL approach yielded an average time savings of 65.6% per case, with variations from 45.95% to 79.56%. Conclusion This study highlights the potential of using LLMs for automated clinical data extraction from medical reports. Incorporating HITL annotations enhances precision and also ensures the reliability of the extracted data. This methodology presents a scalable privacy-preserving option that can significantly support clinical documentation and research endeavors.

The aim of this study was to develop a deep learning-based algorithm for fully automated spleen segmentation using CT images and to evaluate the performance in conditions directly or indirectly affecting the spleen (e.g., splenomegaly, ascites). For this, a 3D U-Net was trained on an in-house dataset (n = 61) including diseases with and without splenic involvement (in-house U-Net), and an open-source dataset from the Medical Segmentation Decathlon (open dataset, n = 61) without splenic abnormalities (open U-Net). Both datasets were split into a training (n = 32.52%), a validation (n = 9.15%) and a testing dataset (n = 20.33%). The segmentation performances of the two models were measured using four established metrics, including the Dice Similarity Coefficient (DSC). On the open test dataset, the in-house and open U-Net achieved a mean DSC of 0.906 and 0.897 respectively (p = 0.526). On the in-house test dataset, the in-house U-Net achieved a mean DSC of 0.941, whereas the open U-Net obtained a mean DSC of 0.648 (p < 0.001), showing very poor segmentation results in patients with abnormalities in or surrounding the spleen. Thus, for reliable, fully automated spleen segmentation in clinical routine, the training dataset of a deep learning-based algorithm should include conditions that directly or indirectly affect the spleen.

Splenomegaly is a common cross-sectional imaging finding with a variety of differential diagnoses. This study aimed to evaluate whether a deep learning model could automatically segment the spleen and identify the cause of splenomegaly in patients with cirrhotic portal hypertension versus patients with lymphoma disease. This retrospective study included 149 patients with splenomegaly on computed tomography (CT) images (77 patients with cirrhotic portal hypertension, 72 patients with lymphoma) who underwent a CT scan between October 2020 and July 2021. The dataset was divided into a training (n = 99), a validation (n = 25) and a test cohort (n = 25). In the first stage, the spleen was automatically segmented using a modified U-Net architecture. In the second stage, the CT images were classified into two groups using a 3D DenseNet to discriminate between the causes of splenomegaly, first using the whole abdominal CT, and second using only the spleen segmentation mask. The classification performances were evaluated using the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE). Occlusion sensitivity maps were applied to the whole abdominal CT images, to illustrate which regions were important for the prediction. When trained on the whole abdominal CT volume, the DenseNet was able to differentiate between the lymphoma and liver cirrhosis in the test cohort with an AUC of 0.88 and an ACC of 0.88. When the model was trained on the spleen segmentation mask, the performance decreased (AUC = 0.81, ACC = 0.76). Our model was able to accurately segment splenomegaly and recognize the underlying cause. Training on whole abdomen scans outperformed training using the segmentation mask. Nonetheless, considering the performance, a broader and more general application to differentiate other causes for splenomegaly is also conceivable.

This study evaluates the diagnostic performance of commercial and open-source Vision-Language Models (VLMs) in neuroradiological image interpretation, using a dataset of 100 brain and spine cases from Radiopaedia. Five VLMs (Gemini 2.0, OpenAI o1, Llama 3.2 90b, Qwen 2.5, Grok-2-vision) were compared to expert neuroradiologists in generating differential diagnoses based on brief clinical presentations and imaging. Neuroradiologists achieved a mean accuracy of 86.2%, whereas the best-performing VLM (Gemini 2.0) reached 35%. Evaluation of the top three differentials improved VLM accuracy marginally, but remained inferior to human experts. Clinical harm analysis revealed frequent diagnostic risks, primarily treatment delays, with harmful outputs in up to 45% of cases. Error analysis showed consistent failure modes including incorrect anatomical localization, inaccurate imaging descriptions, and hallucinated findings. These results highlight the current limitations of VLMs and underscore the importance of expert oversight in neuroradiological diagnosis.