Matthias Guckenberger

Background The rapid growth of biomedical publishing in radiation oncology has created a need for automated methods to extract structured information from unstructured text. Large language models (LLMs) show promise in biomedical natural language processing, but their use for clinical trial data curation remains to be systematically evaluated. Methods We evaluated an LLM-based extraction system for radiation oncology trials on ClinicalTrials.gov. Forty-five classification questions spanning variables such as trial design, tumor type, stage, radiotherapy modality, systemic therapy, and outcomes were implemented using the data-element-extractor Python library with three open-source LLMs (GPT-OSS-20B, GPT-OSS-120B, and Qwen-3.6-27B). A manually curated ground-truth of 94 trials was used for performance evaluation based on accuracy, precision, recall, and F1-score. Results Qwen-3.6-27B achieved the highest overall accuracy of 92.0% across all questions. For the 39 Boolean questions, it demonstrated 93.0% accuracy, 81.6% precision, 83.7% recall, and an F1-score of 82.6%, with 23 Boolean questions achieving an F1-score ≥80.0%. Three numerical questions reached accuracy ≥80.0%. Applied to the full dataset, the system characterized the clinical landscape of radiotherapy trials, detailing the prevalence of advanced techniques (e.g., 13.7% SBRT) and the integration of systemic agents, including chemotherapy (57.3%) and immunotherapy (26.1%). Conclusion LLM-based extraction can enable accurate, scalable curation of radiotherapy trial data. While challenges remain with complex categorical or numerical variables, open-source LLMs can serve as effective components of transparent, auditable clinical research pipelines.

Background and purpose Vendor-provided quality assurance (QA) tools integrated in linear accelerators (linac) promise the improvement of efficiency and reproducibility of QA, but their ability to detect relevant machine deviations has not been comprehensively evaluated in literature. This study evaluated the sensitivity of automated tests to detect deliberately introduced miscalibrations in multiple linac sub-systems and compared their performance with established QA methods. Materials and methods A Varian Accela linac was systematically mistuned in respect to beam output, symmetry, quality, position; in MLC configuration, external motion axes, and isocentricity. The daily required Machine Performance Check (MPC), automated QA tool MaximQA (MQA) and also system calibration (SysCal) were assessed after each introduced deviation. Reference measurements included DailyQA3, Stakitt, IsoLock. Results The automated MPC and QA tools consistently detected introduced beam output and beam symmetry changes, with MQA showing the closest agreement to DailyQA3 for beam characteristics. MLC miscalibrations were reliably detected by MPC, whereas MQA picket fence and Winston-Lutz tests were less sensitive. External axis miscalibrations and isocentricity shifts were detected by both MPC and MQA with good agreement to reference measurements. Conclusions The vendor-provided MPC and QA procedures were sensitive to the introduced clinically relevant deviations on this C-arm linac. MPC outperformed MQA for MLC-related tests, whereas MQA showed better agreement for beam characteristics. Together, these complementary tools support safe daily operation after the required MPC and efficient implementation of periodic machine QA, which was shown to be sensitive to machine deviations.

Objectives: We evaluated the performance of mammography-based radiomics in prediction of clinical nodal status, clinical tumor stage, clinical disease stage, status of the PIK3CA mutation and concordance with the radiologist-assessed BI-RADS category. Mammography is often the first imaging modality performed in the screening or diagnostic workup of breast cancer. Materials and Methods: In this single-center retrospective study, we included 102 histopathologically confirmed cases of breast cancer from 100 patients. The tumor region and whole-breast parenchyma were contoured on 368 craniocaudal and mediolateral-oblique mammograms. Cases were temporally split into training and test sets by histopathological diagnosis date (training: 70%, held-out test: 30%). Six linear models were tuned by 5-fold x 3-repeat stratified cross-validation. The winning model per endpoint was applied once to the test set. We report AUCs with 95% confidence intervals, permutation p-values and Benjamini–Hochberg false discovery rate (BH-FDR) correction across five endpoints. Results: Three of the studied endpoints reached significance after BH-FDR correction: clinical nodal status (cN0 vs. cN-positive; AUC: 0.726 (95% CI: 0.501–0.886)), clinical tumor stage (cT1-2 vs. cT3-4; AUC: 0.792 (95% CI: 0.575–0.929)) and overall disease stage (I–II vs. III–IV; AUC: 0.778 (95% CI: 0.572–0.917)). Mammography-based radiomics failed to predict the presence of PIK3CA mutation and concordance with radiologist-assessed BI-RADS category. Conclusions: Mammography-based radiomics has shown a hypothesis-generating discriminative value for differentiation between cN-negative and cN-positive disease, early from advanced clinical tumor stage, and early from advanced overall disease stage in breast cancer. Mammography-based radiomics did not predict PIK3CA status and did not discriminate between radiologist-assessed BI-RADS 4 and 5 groups.

Background and purpose: MR-guided radiotherapy (MRgRT) offers superior soft-tissue contrast and enables adaptive treatments which improve tumor targeting and organ-at-risk sparing. However, integrated MR-Linac systems remain financially and logistically inaccessible for many hospitals. To address this gap, we implemented a modular MRgRT workflow — termed the “Virtual MR-Linac”— which decouples MR imaging and treatment delivery, enabling daily adaptive radiotherapy (ART) using diagnostic-quality MRI. Materials and methods: Between November 2023 and January 2025, 31 male patients with pelvic lymph node metastases were treated using this workflow. Daily MR scans were acquired in treatment position using a 1.5T MR scanner. Contours were adapted on 2D T2 turbo-spin-echo images, while synthetic CTs generated from T1 DIXON sequences were used for dose calculation. Adapted plans were delivered on a C-arm Linac using CBCT imaging for positioning. Quality assurance included independent Monte Carlo dose recalculations, automated plan parameters verification and a plan complexity check using in-house scripting. Results: Across 99 adapted fractions, plan adaptation improved dosimetric outcomes, with a median 14% increase in Planning Target Volume (PTV) coverage and a 7% reduction in bowel D1cc (dose to 1 cc). Median times for MR imaging, contouring, and planning were 12, 12, and 20 min, respectively. Conclusions: The Virtual MR-Linac approach enables high-quality ART to compensate for inter-fractional motion without the costs and complexity of integrated systems. It improves resource utilization, maintains patient comfort, and offers a flexible, scalable solution for clinical implementation.

Objectives We evaluated the performance of mammography-based radiomics in prediction of clinical nodal status, clinical tumor stage, clinical disease stage, status of the PIK3CA mutation and concordance with the radiologist-assessed BI-RADS category. Mammography is often the first imaging modality performed in the screening or diagnostic workup of breast cancer. Materials & Methods In this single-center retrospective study, we included 102 histopathologically-confirmed cases of breast cancer from 100 patients. The tumor region and whole-breast parenchyma were contoured on 368 craniocaudal and mediolateral-oblique mammograms. Cases were temporally split into training and test sets by histopathological diagnosis date (training 70%, held-out test 30%). Six linear models were tuned by 5-fold x 3-repeat stratified cross-validation. The winning model per endpoint was applied once to the test set. We report AUC with 95% confidence intervals, permutation p-values and Benjamini-Hochberg false discovery rate (BH-FDR) correction across five endpoints. Results Three of the studied endpoints reached significance after BH-FDR correction: clinical nodal status (cN0 vs cN-positive; AUC 0.726 (95% CI: 0.501–0.886)), clinical tumor stage (cT1-2 vs cT3-4; AUC 0.792 (95% CI: 0.575–0.929)) and overall disease stage (I-II vs III-IV; AUC 0.778 (95% CI: 0.572–0.917)). Mammography-based radiomics failed to predict the presence of PIK3CA mutation and concordance with radiologist-assessed BI-RADS category. Conclusion Mammography-based radiomics can be used to discriminate between cN-negative and cN-positive disease, early from advanced clinical tumor stage and early from advanced overall disease stage in breast cancer. Mammography-based radiomics did not predict PIK3CA status and did not discriminate between radiologist-assessed BI-RADS 4 and 5 groups.