S1606
Physics - Autosegmentation
ESTRO 2026
Foundation Trust, Manchester, United Kingdom. 4 Department of Proton Beam Therapy, The Christie Proton Beam Therapy Centre, Manchester, United Kingdom
Results: Our approach significantly reduced annotation time while maintaining sufficient levels of accuracy. Compared to manual and bounding box-based segmentation, our eye-tracking-based method reduced annotation time by over 50% for both observers and yielded an approximately 1.5% gain in accuracy for observer 2 (Table 1). Overall, the eye- tracking-based approach balances segmentation precision and speed, serving as an effective alternative to existing interactive tools.
Purpose/Objective: Dental development is highly sensitive to
chemotherapy and radiotherapy during childhood, leading to long-term side-effects such as dental abnormalities1. Despite this, dose to the teeth is rarely assessed in clinical workflows, as manual contouring is time-consuming and dental maturity varies widely between patients. Advances in deep-learning based tools such as TotalSegmentator2 may enable fully automated and reproducible delineation; however this model was trained on the ToothFairy dataset3, which comprises dental cone-beam CT images from a predominantly adult cohort, raising questions about its performance in primary (milk teeth) and mixed dentitions. This study aimed to evaluate the performance of TotalSegmentator on a paediatric cohort to determine its usefulness in radiotherapy practice and research. Material/Methods: Planning CT scans of five paediatric patients (ages 1, 2, 3, 4 and 12) treated with radiotherapy for head and neck sarcomas were included. A dental specialist manually delineated dental sextants per case as reference contours while automated delineation was performed using TotalSegmentator’s new “Teeth” task3. The resulting 77 dental structures were combined into sextants to match the manual definitions, with minor automated adjustments to close small inter-tooth gaps. Agreement between automated and reference sextants was quantified using mean surface distance (mDTA) and 95th- percentile Hausdorff distance (HD95). Discrepancies were reviewed visually, focusing especially on regions of mixed and primary dentition. Mean absorbed dose per sextant was extracted from the clinical dose distribution for both sets of contours and compared using identity-line plots. Results: TotalSegmentator successfully delineated the dentition in all cases, with no failed or missing structures and a mean mDTA of 0.5 mm (range 0.2-1.3 mm) across all sextants. Small discrepancies were observed around developing dentition (Figure 1), likely due to ambiguity from root development. HD95 values were below 5 mm for 76% of sextants, with outliers up to 12 mm in these regions. Mean absorbed dose per sextant showed strong agreement, with 90% within 2 Gy of the reference, as seen in Figure 2.
Conclusion: Integration of eye-tracking into OARs segmentation shows potential for improving radiotherapy contouring efficiency. In this pilot study, observers guided segmentation using natural visual correction. The eye-tracking-based approach prioritized annotation speed and preserved manual-level accuracy. These findings suggest that eye-tracking interactions may offer a reliable pathway toward more efficient human-AI collaboration in clinical segmentation workflows. References: [1] Chen X et al. "A deep learning-based auto- segmentation system for organs-at-risk on whole-body computed tomography images for radiation therapy." Radiotherapy and Oncology 160 (2021): 175-184.[2] Ma J et al. "Segment anything in medical images." Nature Communications 15.1 (2024): 654.[3] Luo X et al. "WORD: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from CT image." Medical Image Analysis 82 (2022): 102642. Keywords: AI, segmentation, organs-at-risk Digital Poster 4454 Evaluating dental dose in paediatric radiotherapy using automated contouring Thomas B Melichar 1 , Emma Foster-Thomas 2 , Angela Davey 1 , Lucy Davies 1,3 , Eliana M Vasquez Osorio 1 , Shermaine Pan 4 , Marianne C Aznar 1 1 Division of cancer sciences, University of Manchester, Manchester, United Kingdom. 2 University Dental Hospital of Manchester, Manchester University NHS Foundation Trust, Manchester, United Kingdom. 3 Department of Radiotherapy, The Christie NHS
Made with FlippingBook - Share PDF online