S1580
Physics - Autosegmentation
ESTRO 2026
diameters < 0.5 cm. The AI algorithm generated an average of 0.22 FP lesions per patient. Conclusion: This AI-based algorithm shows encouraging sensitivity and DSC performance with a very low FP rate, particularly in larger lesions (> 0.5 cm). Further training with smaller metastases is important in the training of future AI algorithms to improve sensitivity and segmentation accuracy. Keywords: Artificial Intelligence, Automated Segmentation Local Convolutional Dice Similarity (lcDSC): A geometric contour similarity measure with good correlation to visual assessment and clinical usability Aleksandra Zyryanova, Bashar Al-Qaisieh, Michael G Nix Radiotherapy Physics, Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom Purpose/Objective: Common metrics for assessment of autosegmentations against gold-standard [1] (e.g. Dice Similarity (DSC), Mean and Max Distance-to- Digital Poster Highlight 3150 Agreement (MeanDA, MaxDA)) are notoriously poorly correlated [2] to clinician assessment of the need for editing. This leads to reliance on visual assessment for confidence at commissioning and in clinical use, which is subjective, time-consuming and error- prone.We introduce local convolutional DSC (lcDSC), providing a visual map of the regions of discordance and improved correlation to clinician Likert scores. Material/Methods: lcDSC was implemented in Python in the RayStation 2025A-R scripting environment. A spherical 5 voxel kernel was convolved with masks of test and reference ROIs, and their intersection to yield Tv, Rv and Iv respectively. DSC was computed per-voxel as; DSCvox = 2 . | Iv | / (| Tv | + | Rv |)and presented as a visual overlay (figure 1). Only voxels whose kernels contained an ROI boundary in at least one mask were computed. lcDSC was computed as the fraction of these voxels for which DSCvox < 0.01.Sub-regions of DSCvox<0.01 were separated by morphological opening and their connected components were computed. These components were ranked by inverse likelihood, relative to the cohort distribution, to identify outliers. lcDSC_connect incorporated the minimum component likelihood (Lmin) per ROI with lcDSC: lcDSC_connect = min(-log10(Lmin ),lcDSC)Both metrics were compared to clinician Likert scores using Pearson correlation. Autosegmented OARs for 23
Oesophgeal RT patients were used. Gold-standard contours were available from deep-learning segmentation commissioning. Results: Both lcDSC and its connectivity-aware derivative lcDSC_connect showed improved correlation with Likert scores, in comparison to conventional metrics and surface DSC [3] (table 1).
Our new metrics were robust to failure cases for stomach, kidney, liver and lung with R > 0.65, while no other metric was well correlated with Likert scores for all ROIs.
lcDSC_connect further improved the correlation to Likert, especially for stomach and spleen. It also allowed identification of statistically outlying regions. Outlier regions were atypically large (figure 1 - gall bladder inclusion in liver), or distant from the ROI centre-of-mass, or at an unusual angle to the COM,
Made with FlippingBook - Share PDF online