Teaching organ-at-risk delineation in the AI era: educational impact and dosimetric relevance of a structured training intervention

Accurate delineation of organs at risk (OARs) remains a cornerstone of radiotherapy planning. As AI-assisted auto-contouring enters routine practice, the place of contouring in residency training is increasingly debated: should residents still be trained to contour, and if so, with what focus? To evaluate the impact of a structured educational intervention on OAR contouring accuracy among radiation oncology residents, and to assess whether training-related changes in contouring performance are reflected in geometric and dosimetric endpoints, with an independent commercial AI auto-segmentation system used as a descriptive benchmark. In this exploratory single-institution study, ten residents independently contoured OARs on a left breast and regional nodal irradiation case before and seven weeks after a two-hour targeted teaching session, with an eight-week interval between sessions and no control group. Expert-validated reference contours were defined by a senior radiation oncologist following routine clinical practice, consisting of expert review and correction of AI-assisted auto-segmentation outputs (MVision AI™ v1.2.7). For contextual benchmarking, an additional commercial AI auto-segmentation system (RayStation ® v23B) was evaluated descriptively. Geometric accuracy was assessed using the Dice Similarity Coefficient (DSC) and the 95th-percentile Hausdorff Distance (HD95). To assess dosimetric relevance, the same fixed VMAT dose distribution was applied to each contour set, and dose–volume metrics were recalculated without re-optimisation, so that the analysis reflects the effect of contour variation on reported dose–volume values rather than on delivered dose. Training was associated with a significant overall improvement in contouring accuracy in mixed-effects analysis (DSC β = +0.034, 95% CI [+ 0.017 ; +0.052], p < 0.001; HD95 geometric ratio 0.788, 95% CI [0.694 ; 0.896], p < 0.001, i.e. a 21.2% reduction). Exploratory structure-specific analyses showed the largest effects for the brachial plexuses (DSC, rank-biserial r = + 0.93 and + 0.82) and for the left anterior descending coronary artery (HD95, r = + 0.85), although no structure-specific comparison remained significant after adjustment for multiple testing. Dosimetric differences were small, none remained significant after adjustment, and overall compliance with mandatory constraints did not change significantly. Post-training accuracy was broadly comparable to that of the evaluated commercial auto-segmentation system, which performed less well for the small tubular structure available for comparison. A short, structured teaching session was associated with improved OAR contouring accuracy, with the largest changes observed for anatomically complex structures. These findings are exploratory and, in the absence of a control group and with residents reassessed on the same case, should be interpreted as associations rather than as a demonstrated causal effect. The observed changes are potentially relevant to plan evaluation rather than demonstrably clinically meaningful. As automated segmentation becomes routine, whether the ability to critically appraise and correct AI-generated contours requires dedicated instruction, or develops from conventional contouring training, remains an open question that this study was not designed to address.