ESTRO 2025 - Abstract Book

S2551

Physics - Autosegmentation

ESTRO 2025

streamline this task, we have enhanced an AI segmentation model to incorporate user inputs (called prompts) applied across entire 2D slices. This approach aims to enable real-time adjustments and accelerate the generation of clinically acceptable GTVs contours 1,2,3.

Material/Methods: For this study, we selected 20 anonymized CT scans featuring lung tumors of varying sizes (mean volume 97,57 cc) and locations. First, three radiation oncologists (ROs) manually delineated the GTVs for each scan, while recording the time required. Then, the same ROs were tasked to delineate the same GTVs using the "Active Contouring" feature—a novel tool developed by Synaptiq—and document the time spent, as well as the number of AI prompts required to achieve clinically acceptable lung GTVs contours. "Active Contouring" employs an advanced deep neural network that incorporates 2D input feedback from ROs to adjust the 3D segmented volumes dynamically, responding to the manual input adjustments within seconds (~1-2s). Results: The results of this study demonstrate a significant improvement in efficiency when using Synaptiq's "Active Contouring" tool for Gross Tumor Volume (GTV) delineation compared to manual methods. The mean time required for delineation with the AI-assisted tool (GTV_A) was 29,12 seconds, whereas manual delineation (GTV_M) required an average of 280.91 seconds. Overall, the "Active Contouring" tool achieved an impressive mean reduction of 89,63% for delineation time (p=0.049). The number of required interactions with the AI-assisted tool was on average 3.58 prompts per volume (range 1–8). This suggest that while some cases required more manual input, the tool was generally comparable with manual delineation in terms of adjustments. The DICE score between each of the 69 GTV_M and GTV_A, was calculated, with a mean score of 0,900 [0,706 - 0,966].

Conclusion: Our study shows that "Active Contouring," enhanced by user input, provides a faster, more reliable alternative to manual GTV delineation. Real-time adjustments to AI-generated contours based on clinician feedback significantly cut the time ROs need to achieve clinically acceptable results, streamlining radiotherapy planning workflows.

Keywords: GTV, prompt, autosegmentation

References: 1. Vaassen F, et al. Real-world analysis of manual editing of deep learning contouring in the thorax region. Phys Imaging Radiat Oncol. 2022;22:104-110. Published 2022 May 14. doi:10.1016/j.phro.2022.04.008 2. A., Mehta, et al. "DeepEdit: deep editable learning for interactive segmentation of 3D medical images," in MICCAI Workshop on Data Augmentation, Labelling, and Imperfections, 2022, pp. 11-21, doi:10.1007/978-3-031-17027-0_2.