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.

Facial recognition technology in medical imaging, particularly with head scans, poses privacy risks due to identifiable facial features. This study evaluates the use of facial recognition software in identifying facial features from head CT scans and explores a defacing pipeline using TotalSegmentator to reduce re-identification risks while preserving data integrity for research. 1404 high-quality renderings from the UCLH EIT Stroke dataset, both with and without defacing were analysed. The performance of defacing with the face mask created by TotalSegmentator was compared to a state-of-the-art CT defacing algorithm. Face detection was performed using deep learning models. The cosine similarity between facial embeddings for intra- and inter-patient images was compared. A Support Vector Machine was trained on cosine similarity values to assess defacing performance, determining if two renderings came from the same patient. This analysis was conducted on defaced and non-defaced images using 5-fold cross-validation. Faces were detected in 76.5 % of non-defaced images. Intra-patient images exhibited a median cosine similarity of 0.65 (IQR: 0.47-0.80), compared to 0.50 (IQR: 0.39-0.62) for inter-patient images. A binary classifier performed moderately on non-defaced images, achieving a ROC-AUC of 0.69 (SD = 0.01) and an accuracy of 0.65 (SD = 0.01) in distinguishing whether a scan belonged to the same or a different individual. Following defacing, performance declined markedly. Defacing with the TotalSegmentator decreased the ROC-AUC to 0.55 (SD = 0.02) and the accuracy to 0.56 (SD = 0.01), whereas the CTA-DEFACE algorithm brought the performance down to a ROC-AUC of 0.60 (SD = 0.02) and an accuracy of 0.59 (SD = 0.01). These results demonstrate the effectiveness of defacing algorithms in mitigating re-identification risks, with the TotalSegmentator providing slightly superior privacy protection. Facial recognition software can identify facial features from partial and complete head CT scan renderings. However, using the TotalSegmentator to deface images reduces re-identification risks to a near-chance level. We offer code to implement this privacy-preserving pipeline. Utilizing the TotalSegmentator framework, the proposed pipeline efficiently removes facial features from CT images, making it ideal for multi-site research and data sharing. It is a useful tool for radiographers and radiologists who must comply with medico-legal requirements necessitating the removal of facial features.

Intracerebral hemorrhage (ICH) associated with primary and metastatic brain tumors presents a significant challenge in neuro-oncology due to the substantial risk of complications