The manuscript proposes employing the nnU-Net architecture for the segmentation of the aorta from computed tomography images. Preprocessing involved transforming various value ranges to align with the Hounsfield unit scale, in addition to implementing a windowing procedure. During the network training phase, a five-fold cross-validation protocol was utilized, resulting in the training of five models, whose predictions on the test set were aggregated. The models achieved an average Dice coefficient of 0.921 and an Intersection over Union of 0.858, signifying high prediction quality. The value of the Hausdorff distance indicated the presence of outliers, while the analysis of three-dimensional reconstructions highlighted regions of diminished segmentation accuracy.

Purpose: Artificial intelligence is increasingly developed and deployed to support radiological image interpretation, triage, segmentation and workflow prioritisation. However, high performance reported in retrospective studies does not necessarily translate into safe and useful deployment in routine practice. This narrative review examines why this translational gap persists and what radiology departments, researchers and developers should address before clinical implementation. Methods: We synthesised methodological, empirical and regulatory literature on the development, validation, deployment and governance of radiology AI. Radiology was used as the primary setting, with adjacent imaging fields discussed only where they illustrate broader mechanisms relevant to medical image analysis. Results: Recurrent barriers include limited data quality and representativeness, scanner- and protocol-related domain shift, shortcut learning, hidden stratification, insufficient external and prospective validation, poor calibration, over-reliance on global metrics and weak workflow integration. Model performance can be shaped by scanner protocols, reconstruction methods, local reporting practices, patient selection and institutional workflows beyond the underlying pathology. In Europe, GDPR-based data governance, the Medical Device Regulation, the AI Act and the European Health Data Space further shape clinical deployment. Conclusions: Emerging approaches, including foundation models, generative AI, multimodal systems, federated learning and local adaptation, may address selected technical constraints, but they introduce new uncertainties and do not replace independent validation in the intended clinical setting. Radiology AI should be evaluated not only as an algorithmic model, but as a clinical decision-support component embedded in radiology practice, requiring external and preferably prospective validation, subgroup analysis, calibration assessment, human oversight, workflow integration and post-deployment monitoring.