S1581
Physics - Autosegmentation
ESTRO 2026
indicating they were in a part of the segmentation which was normally robust across the cohort. Conclusion: lcDSC is a novel metric for ROI comparison, with good correlation to clinician Likert scores, enabling semi- automated evaluation of autosegmentation, and providing error maps for visualisation. lcDSC_connect can statistically identify outlier regions within an lcDSC map, enabling interactive assessment of failure cases at commissioning or ongoing monitoring of deep- learning based autosegmentation systems. References: [1] Sherer M et. al. Metrics to evaluate the performance of auto-segmentation for radiation treatment planning: A critical review, Radiother Oncol, 2021 Sep; 10.1016/j.radonc.2021.05.003[2] Podobnik G et al. Geometric, dosimetric and psychometric evaluation of three commercial AI software solutions for OAR auto-segmentation in head and neck radiotherapy. Scientific Reports, 2025 Sep, 10.1038/s41598-025-18598-3[3] Nikolov, S et. al. arXiv:1809.04430 [cs.CV] 10.48550/arXiv.1809.04430 Keywords: similarity metric, commissioning, clinical utility
obtained with the two AI tools, were compared with their MC counterparts, using the quantitative metrics volumetric/surface Dice similarity coefficients (vDSC/sDSC) and maximum/average Hausdorff distance (HD/aHD). The constructed AI-CTVs and MC counterparts were also blinded and evaluated by two radiotherapists using two qualitative methods: (1) a “Grading test,” employing a five-point Likert scale to assess overall quality and estimated time required to correct compared to MC from scratch; and (2) a “Turing test,” involving a preference choice between the AI-CTV and MC-CTV in selected slices, with the median of the two reviewers’ results used for the analyses.
Digital Poster 3399
Qualitative and quantitative evaluation of auto- segmented clinical target volumes and organs at risk in radiotherapy of rectal cancer Albert Siegbahn 1,2 , Love Dahlstedt-Hassler 3 , Rafat Kojoj 2 , Cecilia Lagerbäck 3 , Anna Schedin 3 , Pehr Lind 1,3 1 Department of Clinical Science and Education, Karolinska Institutet, Stockholm, Sweden. 2 Department of Medical Physics, Södersjukhuset, Stockholm, Sweden. 3 Department of Oncology, Södersjukhuset, Stockholm, Sweden Purpose/Objective: Manual contouring (MC) is a time-consuming task in rectal-cancer radiotherapy (RT) planning. Artificial intelligence (AI) can reduce the time required to delineate the clinical target volumes (CTVs) and organs of interest (OOIs). In this study, we evaluated the quality of auto-segmented OOIs and the constructed CTVs based on the available auto-segmentations. Material/Methods: Dose planning data were collected from ten patients who underwent preoperative RT for locally advanced rectal cancer in 2024. Auto-segmented structures from the AI-Rad and MVision software tools were incorporated. We constructed AI-CTVs based on MVision segmentations in order to compare them to the MC-CTVs, in a step-by-step method with minimal manual segmentation described in Fig. 1. Also, AI- OOIs, i.e. bladder, femoral heads, and bowel bag,
Results: The mean vDSC, sDSC, HD, and aHD values of our constructed AI-CTVs with the MC-CTVs used as reference were 0.86, 0.65, 23.4 mm, and 0.56 mm, respectively. For both AI tools, the agreement in the OOI metrics was overall good, although it was less consistent for the bowel bag.In both qualitative evaluations, comparing the AI-CTVs with the MC-CTVs (Table 1), the MC-CTVs were clearly preferred, with the AI-CTVs requiring major revisions. The cranial-anterior nodal levels (anatomical volumes) had poorer coverage, reflecting differences between the contouring guidelines used to obtain the AI training set and those used in local clinical practice.
Made with FlippingBook - Share PDF online