Uzbekistan has mandated hybrid artificial intelligence (AI) and radiologist reading of screening mammograms, but no AI system has been clinically evaluated in its population. The aim of the study was to assess the diagnostic performance of a deep learning model for breast cancer detection on mammograms from Uzbekistan and to compare it with the primary radiologist. Object and methods. Retrospective multicentre diagnostic accuracy study of 1,759 anonymised full-field digital mammography examinations with a verified outcome, selected from 33,599 examinations performed from March 2022 to April 2023 in Tashkent, Nukus and Samarkand. A YOLO11-based detection model was trained only on public datasets (Radiological Society of North America (RSNA) and VinDr-Mammo); the Uzbek data were used solely for testing. The reference standard was histopathology, clinical confirmation or 6–24-month imaging follow-up. The comparator was the primary radiologist’s Breast Imaging Reporting and Data System (BI-RADS) 4–5 assessment (biopsy recommendation); alternative handling of incomplete (BI-RADS 0) assessments was tested in sensitivity analyses. ROC analysis, bootstrap confidence intervals (CI), McNemar’s test and multivariable logistic regression were used. Results. Breast cancer was confirmed in 316 women (18.0%). The model achieved an area under the curve (AUC) of 0.914 (95% confidence interval — CI 0.895–0.932); at the Youden threshold (0.495) sensitivity was 0.829 and specificity 0.852, versus 0.968 and 0.836 for the radiologist. Overall accuracy did not differ significantly (McNemar p=0.31), but the radiologist was more sensitive (p <0.001). When the 70 BI-RADS 0 examinations (4 cancers, all scored above the AI threshold) were counted as positive, the radiologist’s specificity fell to 0.791, the AI became significantly more specific (p <0.001) and overall accuracy still did not differ (p=0.07). All 5 interval cancers missed by the radiologist (BI-RADS 1–2) received AI scores above the threshold. Family history, hormone therapy, age, body mass index (BMI), previous biopsy and breast density were independent predictors of cancer. Conclusions. A deep learning model trained on public data showed high discriminative ability on Uzbek mammograms, with overall accuracy comparable to the radiologist but lower sensitivity; it is best suited to the role of a second reader within the national hybrid reading workflow, and prospective validation is required.
Aims: The APPROACH (Analysis of Proton vs Photon Radiotherapy in Oligodendroglioma and Assessment of Cognitive Health) trial is a UK randomised controlled study comparing photon radiotherapy (RT) with proton beam therapy (PBT) for oligodendroglioma, with a primary endpoint of neurocognitive function at five years. As part of the trial's radiotherapy quality assurance (RTQA) programme, a national preaccrual facility questionnaire (FQ) was undertaken to evaluate RT practices across England and Wales and assess readiness for protocol compliance. This study reports UK photon radiotherapy practice for oligodendroglioma and evaluates consistency relative to the APPROACH protocol requirements. Materials and methods: Between December 2022 and September 2025, photon radiotherapy centres in England and Wales (n = 23) and both UK proton centres completed the FQ. The questionnaire captured routine practice in immobilisation, imaging, contouring, planning, and verification. Responses were collated by the National Radiotherapy Trials Quality Assurance (RTTQA) group. Results: All centres routinely used thermoplastic mask immobilisation and delivered photon radiotherapy using intensity-modulated radiation therapy/ volumetric modulated arc therapy (IMRT/VMAT). Dose schedules were consistent, most commonly 54 Gy in 28-30 fractions for World Health Organization (WHO) grade 2 oligodendroglioma and 59.4 Gy in 33 fractions for grade 3.However, substantial variation was identified in pretreatment imaging: Computed tomography (CT) slice thickness ranged from 1 to 3 mm and magnetic resonance imaging (MRI) T1 slice thickness from 0.8 to 3 mm, with nine centres exceeding the protocol-required 1 mm slice thickness. MRI type and sequencing were inconsistently reported, and not all centres used dedicated planning MRI scans. Marked variability was also observed in organ-at-risk (OAR) contouring. While optic pathway structures and the brainstem were routinely delineated, only 4/23 (17%) centres contoured the hippocampi, and none contoured the hypothalamus as part of standard practice. Auto-contouring tools were used by 65% of centres. Conclusion: The findings indicate broad alignment in technique and dose delivery but highlight variability in imaging and OAR delineation. The implementation of the APPROACH trial and associated RTQA processes is hoped to promote national standards in oligodendroglioma RT practice. (c) 2026 The Royal College of Radiologists. Published by Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
