Search

Objectives: Structured reporting enhances comparability, readability, and content detail. Large language models (LLMs) could convert free text into structured data without disrupting radiologists' reporting workflow. This study evaluated an on-premise, privacy-preserving LLM for automatically structuring free-text radiology reports., Materials and Methods: We developed an approach to controlling the LLM output, ensuring the validity and completeness of structured reports produced by a locally hosted Llama-2-70B-chat model. A dataset with de-identified narrative chest radiograph (CXR) reports was compiled retrospectively. It included 202 English reports from a publicly available MIMIC-CXR dataset and 197 German reports from our university hospital. Senior radiologist prepared a detailed, fully structured reporting template with 48 question-answer pairs. All reports were independently structured by the LLM and two human readers. Bayesian inference (Markov chain Monte Carlo sampling) was used to estimate the distributions of Matthews correlation coefficient (MCC), with [-0.05, 0.05] as the region of practical equivalence (ROPE)., Results: The LLM generated valid structured reports in all cases, achieving an average MCC of 0.75 (94% HDI: 0.70-0.80) and F1 score of 0.70 (0.70-0.80) for English, and 0.66 (0.62-0.70) and 0.68 (0.64-0.72) for German reports, respectively. The MCC differences between LLM and humans were within ROPE for both languages: 0.01 (-0.05 to 0.07), 0.01 (-0.05 to 0.07) for English, and -0.01 (-0.07 to 0.05), 0.00 (-0.06 to 0.06) for German, indicating approximately comparable performance., Conclusion: Locally hosted, open-source LLMs can automatically structure free-text radiology reports with approximately human accuracy. However, the understanding of semantics varied across languages and imaging findings., Key Points: Question Why has structured reporting not been widely adopted in radiology despite clear benefits and how can we improve this? Findings A locally hosted large language model successfully structured narrative reports, showing variation between languages and findings. Critical relevance Structured reporting provides many benefits, but its integration into the clinical routine is limited. Automating the extraction of structured information from radiology reports enables the capture of structured data while allowing the radiologist to maintain their reporting workflow., (© 2024. The Author(s).)

Objectives: To investigate the model-, code-, and data-sharing practices in the current radiomics research landscape and to introduce a radiomics research database., Methods: A total of 1254 articles published between January 1, 2021, and December 31, 2022, in leading radiology journals (European Radiology, European Journal of Radiology, Radiology, Radiology: Artificial Intelligence, Radiology: Cardiothoracic Imaging, Radiology: Imaging Cancer) were retrospectively screened, and 257 original research articles were included in this study. The categorical variables were compared using Fisher's exact tests or chi-square test and numerical variables using Student's t test with relation to the year of publication., Results: Half of the articles (128 of 257) shared the model by either including the final model formula or reporting the coefficients of selected radiomics features. A total of 73 (28%) models were validated on an external independent dataset. Only 16 (6%) articles shared the data or used publicly available open datasets. Similarly, only 20 (7%) of the articles shared the code. A total of 7 (3%) articles both shared code and data. All collected data in this study is presented in a radiomics research database (RadBase) and could be accessed at https://github.com/EuSoMII/RadBase ., Conclusion: According to the results of this study, the majority of published radiomics models were not technically reproducible since they shared neither model nor code and data. There is still room for improvement in carrying out reproducible and open research in the field of radiomics., Clinical Relevance Statement: To date, the reproducibility of radiomics research and open science practices within the radiomics research community are still very low. Ensuring reproducible radiomics research with model-, code-, and data-sharing practices will facilitate faster clinical translation., Key Points: • There is a discrepancy between the number of published radiomics papers and the clinical implementation of these published radiomics models. • The main obstacle to clinical implementation is the lack of model-, code-, and data-sharing practices. • In order to translate radiomics research into clinical practice, the radiomics research community should adopt open science practices., (© 2023. The Author(s).)

Objective: To conduct a comprehensive bibliometric analysis of artificial intelligence (AI) and its subfields as well as radiomics in Radiology, Nuclear Medicine, and Medical Imaging (RNMMI)., Methods: Web of Science was queried for relevant publications in RNMMI and medicine along with their associated data from 2000 to 2021. Bibliometric techniques utilised were co-occurrence, co-authorship, citation burst, and thematic evolution analyses. Growth rate and doubling time were also estimated using log-linear regression analyses., Results: According to the number of publications, RNMMI (11,209; 19.8%) was the most prominent category in medicine (56,734). USA (44.6%) and China (23.1%) were the two most productive and collaborative countries. USA and Germany experienced the strongest citation bursts. Thematic evolution has recently exhibited a significant shift toward deep learning. In all analyses, the annual number of publications and citations demonstrated exponential growth, with deep learning-based publications exhibiting the most prominent growth pattern. Estimated continuous growth rate, annual growth rate, and doubling time of the AI and machine learning publications in RNMMI were 26.1% (95% confidence interval [CI], 12.0-40.2%), 29.8% (95% CI, 12.7-49.5%), and 2.7 years (95% CI, 1.7-5.8), respectively. In the sensitivity analysis using data from the last 5 and 10 years, these estimates ranged from 47.6 to 51.1%, 61.0 to 66.7%, and 1.4 to 1.5 years., Conclusion: This study provides an overview of AI and radiomics research conducted mainly in RNMMI. These results may assist researchers, practitioners, policymakers, and organisations in gaining a better understanding of both the evolution of these fields and the importance of supporting (e.g., financial) these research activities., Key Points: • In terms of the number of publications on AI and ML, Radiology, Nuclear Medicine, and Medical Imaging was the most prominent category compared to the other categories related to medicine (e.g., Health Policy & Services, Surgery). • All evaluated analyses (i.e., AI, its subfields, and radiomics), based on the annual number of publications and citations, demonstrated exponential growth, with decreasing doubling time, which indicates increasing interest from researchers, journals, and, in turn, the medical imaging community. • The most prominent growth pattern was observed in deep learning-based publications. However, the further thematic analysis demonstrated that deep learning has been underdeveloped but highly relevant to the medical imaging community., (© 2023. The Author(s), under exclusive licence to European Society of Radiology.)

Objectives: To compare the accuracy of lesion detection of trauma-related injuries using combined "all-in-one" fused (AIO) and conventionally reconstructed images (CR) in acute trauma CT., Methods: In this retrospective study, trauma CT of 66 patients (median age 47 years, range 18-96 years; 20 female (30.3%)) were read using AIO and CR. Images were independently reviewed by 4 blinded radiologists (two residents and two consultants) for trauma-related injuries in 22 regions. Sub-analyses were performed to analyze the influence of experience (residents vs. consultants) and body region (chest, abdomen, skeletal structures) on lesion detection. Paired t-test was used to compare the accuracy of lesion detection. The effect size was calculated (Cohen's d). Linear mixed-effects model with patients as the fixed effect and random forest models were used to investigate the effect of experience, reconstruction/image processing, and body region on lesion detection., Results: Reading time of residents was significantly faster using AIO (AIO: 266 ± 72 s, CR: 318 ± 113 s; p < 0.001; d = 0.46) while no significant difference was observed in the accuracy of lesion detection (AIO: 93.5 ± 6.0%, CR: 94.6 ± 6.0% p = 0.092; d =  - 0.21). Reading time of consultants showed no significant difference (AIO: 283 ± 82 s, CR: 274 ± 95 s; p = 0.067; d = 0.16). Accuracy was significantly higher using CR; however, the difference and effect size were very small (AIO 95.1 ± 4.9%, CR: 97.3 ± 3.7%, p = 0.002; d =  - 0.39). The linear mixed-effects model showed only minor effect of image processing/reconstruction for lesion detection., Conclusions: Residents at the emergency department might benefit from faster reading time without sacrificing lesion detection rate using AIO for trauma CT., Key Points: • Image fusion techniques decrease the reading time of acute trauma CT without sacrificing diagnostic accuracy., (© 2022. The Author(s), under exclusive licence to European Society of Radiology.)

Objectives: To investigate the robustness of radiomic features between three dual-energy CT (DECT) systems., Methods: An anthropomorphic body phantom was scanned on three different DECT scanners, a dual-source (dsDECT), a rapid kV-switching (rsDECT), and a dual-layer detector DECT (dlDECT). Twenty-four patients who underwent abdominal DECT examinations on each of the scanner types during clinical follow-up were retrospectively included (n = 72 examinations). Radiomic features were extracted after standardized image processing, following ROI placement in phantom tissues and healthy appearing hepatic, splenic and muscular tissue of patients using virtual monoenergetic images at 65 keV (VMI 65keV ) and virtual unenhanced images (VUE). In total, 774 radiomic features were extracted including 86 original features and 8 wavelet transformations hereof. Concordance correlation coefficients (CCC) and analysis of variances (ANOVA) were calculated to determine inter-scanner robustness of radiomic features with a CCC of ≥ 0.9 deeming a feature robust., Results: None of the phantom-derived features attained the threshold for high feature robustness for any inter-scanner comparison. The proportion of robust features obtained from patients scanned on all three scanners was low both in VMI 65keV (dsDECT vs. rsDECT:16.1% (125/774), dlDECT vs. rsDECT:2.5% (19/774), dsDECT vs. dlDECT:2.6% (20/774)) and VUE (dsDECT vs. rsDECT:11.1% (86/774), dlDECT vs. rsDECT:2.8% (22/774), dsDECT vs. dlDECT:2.7% (21/774)). The proportion of features without significant differences as per ANOVA was higher both in patients (51.4-71.1%) and in the phantom (60.6-73.4%)., Conclusions: The robustness of radiomic features across different DECT scanners in patients was low and the few robust patient-derived features were not reflected in the phantom experiment. Future efforts should aim to improve the cross-platform generalizability of DECT-derived radiomics., Key Points: • Inter-scanner robustness of dual-energy CT-derived radiomic features was on a low level in patients who underwent clinical examinations on three DECT platforms. • The few robust patient-derived features were not confirmed in our phantom experiment. • Limited inter-scanner robustness of dual-energy CT derived radiomic features may impact the generalizability of models built with features from one particular dual-energy CT scanner type., (© 2021. European Society of Radiology.)

Objectives: To investigate, in patients with metastatic prostate cancer, whether radiomics of computed tomography (CT) image data enables the differentiation of bone metastases not visible on CT from unaffected bone using 68  Ga-PSMA PET imaging as reference standard., Methods: In this IRB-approved retrospective study, 67 patients (mean age 71 ± 7 years; range: 55-84 years) showing a total of 205 68  Ga-PSMA-positive prostate cancer bone metastases in the thoraco-lumbar spine and pelvic bone being invisible in CT were included. Metastases and 86 68  Ga-PSMA-negative bone volumes in the same body region were segmented and further post-processed. Intra- and inter-reader reproducibility was assessed, with ICCs < 0.90 being considered non-reproducible. To account for imbalances in the dataset, data augmentation was performed to achieve improved class balance and to avoid model overfitting. The dataset was split into training, test, and validation set. After a multi-step dimension reduction process and feature selection process, the 11 most important and independent features were selected for statistical analyses., Results: A gradient-boosted tree was trained on the selected 11 radiomic features in order to classify patients' bones into bone metastasis and normal bone using the training dataset. This trained model achieved a classification accuracy of 0.85 (95% confidence interval [CI]: 0.76-0.92, p < .001) with 78% sensitivity and 93% specificity. The tuned model was applied on the original, non-augmented dataset resulting in a classification accuracy of 0.90 (95% CI: 0.82-0.98) with 91% sensitivity and 88% specificity., Conclusion: Our proof-of-concept study indicates that radiomics may accurately differentiate unaffected bone from metastatic bone, being invisible by the human eye on CT., Key Points: • This proof-of-concept study showed that radiomics applied on CT images may accurately differentiate between bone metastases and metastatic-free bone in patients with prostate cancer. • Future promising applications include automatic bone segmentation, followed by a radiomics classifier, allowing for a screening-like approach in the detection of bone metastases., (© 2021. Crown.)

Machine learning offers great opportunities to streamline and improve clinical care from the perspective of cardiac imagers, patients, and the industry and is a very active scientific research field. In light of these advances, the European Society of Cardiovascular Radiology (ESCR), a non-profit medical society dedicated to advancing cardiovascular radiology, has assembled a position statement regarding the use of machine learning (ML) in cardiovascular imaging. The purpose of this statement is to provide guidance on requirements for successful development and implementation of ML applications in cardiovascular imaging. In particular, recommendations on how to adequately design ML studies and how to report and interpret their results are provided. Finally, we identify opportunities and challenges ahead. While the focus of this position statement is ML development in cardiovascular imaging, most considerations are relevant to ML in radiology in general. KEY POINTS: • Development and clinical implementation of machine learning in cardiovascular imaging is a multidisciplinary pursuit. • Based on existing study quality standard frameworks such as SPIRIT and STARD, we propose a list of quality criteria for ML studies in radiology. • The cardiovascular imaging research community should strive for the compilation of multicenter datasets for the development, evaluation, and benchmarking of ML algorithms.

Objectives: The goal of the present study was to classify the most common types of plain radiographs using a neural network and to validate the network's performance on internal and external data. Such a network could help improve various radiological workflows., Methods: All radiographs from the year 2017 (n = 71,274) acquired at our institution were retrieved from the PACS. The 30 largest categories (n = 58,219, 81.7% of all radiographs performed in 2017) were used to develop and validate a neural network (MobileNet v1.0) using transfer learning. Image categories were extracted from DICOM metadata (study and image description) and mapped to the WHO manual of diagnostic imaging. As an independent, external validation set, we used images from other institutions that had been stored in our PACS (n = 5324)., Results: In the internal validation, the overall accuracy of the model was 90.3% (95%CI: 89.2-91.3%), whereas, for the external validation set, the overall accuracy was 94.0% (95%CI: 93.3-94.6%)., Conclusions: Using data from one single institution, we were able to classify the most common categories of radiographs with a neural network. The network showed good generalizability on the external validation set and could be used to automatically organize a PACS, preselect radiographs so that they can be routed to more specialized networks for abnormality detection or help with other parts of the radiological workflow (e.g., automated hanging protocols; check if ordered image and performed image are the same). The final AI algorithm is publicly available for evaluation and extension., Key Points: • Data from one single institution can be used to train a neural network for the correct detection of the 30 most common categories of plain radiographs. • The trained model achieved a high accuracy for the majority of categories and showed good generalizability to images from other institutions. • The neural network is made publicly available and can be used to automatically organize a PACS or to preselect radiographs so that they can be routed to more specialized neural networks for abnormality detection.

Objectives: Interventional radiology (IR) is a growing field but is underrepresented in most medical school curricula. We tested whether endovascular simulator training improves medical students' attitudes towards IR., Materials and Methods: We conducted this prospective study at two university medical centers; overall, 305 fourth-year medical students completed a 90-min IR course. The class consisted of theoretical and practical parts involving endovascular simulators. Students completed questionnaires before the course, after the theoretical and after the practical part. On a 7-point Likert scale, they rated their interest in IR, knowledge of IR, attractiveness of IR, and the likelihood to choose IR as subspecialty. We used a crossover design to prevent position-effect bias., Results: The seminar/simulator parts led to the improvement for all items compared with baseline: interest in IR (pre-course 5.2 vs. post-seminar/post-simulator 5.5/5.7), knowledge of IR (pre-course 2.7 vs. post-seminar/post-simulator 5.1/5.4), attractiveness of IR (pre-course 4.6 vs. post-seminar/post-simulator 4.8/5.0), and the likelihood of choosing IR as a subspecialty (pre-course 3.3 vs. post-seminar/post-simulator 3.8/4.1). Effect was significantly stronger for simulator training compared with that for seminar for all items (p < 0.05). For simulator training, subgroup analysis of students with pre-existing positive attitude showed considerable improvement regarding "interest in IR" (× 1.4), "knowledge of IR" (× 23), "attractiveness of IR" (× 2), and "likelihood to choose IR" (× 3.2) compared with pretest., Conclusion: Endovascular simulator training significantly improves students' attitude towards IR regarding all items. Implementing such courses at a very early stage in the curriculum should be the first step to expose medical students to IR and push for IR., Key Points: • Dedicated IR-courses have a significant positive effect on students' attitudes towards IR. • Simulator training is superior to a theoretical seminar in positively influencing students' attitudes towards IR. • Implementing dedicated IR courses in medical school might ease recruitment problems in the field.

Objectives: To evaluate whether a computed tomography (CT) radiomics-based machine learning classifier can predict histopathology of lymph nodes (LNs) after post-chemotherapy LN dissection (pcRPLND) in patients with metastatic non-seminomatous testicular germ cell tumors (NSTGCTs)., Methods: Eighty patients with retroperitoneal LN metastases and contrast-enhanced CT were included into this retrospective study. Resected LNs were histopathologically classified into "benign" (necrosis/fibrosis) or "malignant" (viable tumor/teratoma). On CT imaging, 204 corresponding LNs were segmented and 97 radiomic features per LN were extracted after standardized image processing. The dataset was split into training, test, and validation sets. After stepwise feature reduction based on reproducibility, variable importance, and correlation analyses, a gradient-boosted tree was trained and tuned on the selected most important features using the training and test datasets. Model validation was performed on the independent validation dataset., Results: The trained machine learning classifier achieved a classification accuracy of 0.81 in the validation dataset with a misclassification of 8 of 36 benign LNs as malignant and 4 of 25 malignant LNs as benign (sensitivity 88%, specificity 72%, negative predictive value 88%). In contrast, a model containing only the LN volume resulted in a classification accuracy of 0.68 with 64% sensitivity and 68% specificity., Conclusions: CT radiomics represents an exciting new tool for improved prediction of the presence of malignant histopathology in retroperitoneal LN metastases from NSTGCTs, aiming at reducing overtreatment in this group of young patients. Thus, the presented approach should be combined with established clinical biomarkers and further validated in larger, prospective clinical trials., Key Points: • Patients with metastatic non-seminomatous testicular germ cell tumors undergoing post-chemotherapy retroperitoneal lymph node dissection of residual lesions show overtreatment in up to 50%. • We assessed whether a CT radiomics-based machine learning classifier can predict histopathology of lymph nodes after post-chemotherapy lymph node dissection. • The trained machine learning classifier achieved a classification accuracy of 0.81 in the validation dataset with a sensitivity of 88% and a specificity of 78%, thus allowing for prediction of the presence of viable tumor or teratoma in retroperitoneal lymph node metastases.

Objectives: To re-evaluate a recently suggested approach of quantifying myocardial oedema and increased tissue inhomogeneity in myocarditis by T2-mapping., Methods: Cardiac magnetic resonance data of 99 patients with myocarditis were retrospectively analysed. Thirthy healthy volunteers served as controls. T2-mapping data were acquired at 1.5 T using a gradient-spin-echo T2-mapping sequence. T2-maps were segmented according to the 16-segments AHA-model. Segmental T2-values, segmental pixel-standard deviation (SD) and the derived parameters maxT2, maxSD and madSD were analysed and compared to the established Lake Louise criteria (LLC)., Results: A re-estimation of logistic regression models revealed that all models containing an SD-parameter were superior to any model containing global myocardial T2. Using a combined cut-off of 1.8 ms for madSD + 68 ms for maxT2 resulted in a diagnostic sensitivity of 75% and specificity of 80% and showed a similar diagnostic performance compared to LLC in receiver-operating-curve analyses. Combining madSD, maxT2 and late gadolinium enhancement (LGE) in a model resulted in a superior diagnostic performance compared to LLC (sensitivity 93%, specificity 83%)., Conclusions: The results show that the novel T2-mapping-derived parameters exhibit an additional diagnostic value over LGE with the inherent potential to overcome the current limitations of T2-mapping., Key Points: • A novel quantitative approach to myocardial oedema imaging in myocarditis was re-evaluated. • The T2-mapping-derived parameters maxT2 and madSD were compared to traditional Lake-Louise criteria. • Using maxT2 and madSD with dedicated cut-offs performs similarly to Lake-Louise criteria. • Adding maxT2 and madSD to LGE results in further increased diagnostic performance. • This novel approach has the potential to overcome the limitations of T2-mapping.