ESTRO 2024 - Abstract Book

S3124

Physics - Autosegmentation

ESTRO 2024

A mono-institutional consecutive series of 40 EBRT-CC and 40 SBRT-PC structure sets (SS) were retrospectively selected. For each anatomical site, 20 CC and 20 PC SS (testing) were auto-segmented by ST, RF, DL1, and DL2 and compared, setting manual contouring as the Gold-Standard. AS times have been registered. Performances have been tested by comparing the resulting 200 sets by means of the Dice Similarity Coefficient (DSC), the Hausdorff Distance (HD), and the Distance-to-Agreement Portion (DAP), i.e. the percentage of automated contours standing 1 mm-, 3 mm-, and 5 mm-far from the manual Gold-Standard (DAP 1mm , DAP 3mm , DAP 5mm ). AS correction times needed by a PC expert radiation oncologist (PC-RO) and 2 CC-expert radiation oncologists (CC-RO1 and CC-RO2) have been collected and a qualitative evaluation has been performed by means of a score table built with 5 degrees of scoring (good agreement, minor differences, edits required, moderate edits required, gross errors) to assess AS clinical validation. The Wilcoxon-Mann-Whitney test has been performed to assess statistical significance (p < 0.05). The median CC AS time per structure set was 21.9 min, 20.9 min, 0.7 min, and 1.1 min for ST, RF, MV, and LI, respectively. PC AS time was 1.8 min and 2.0 min in MV and LI, respectively, and 28.7 and 22.0 in ST and RF. In CC, rectum and bladder DL results showed a significantly increased DSC (e.g., MV vs ST: DSC rectum +34.4%, DSC bladder +20.2%). In PC, DL registered a smaller and not statistically significant DSC increase (e.g., MV vs RF: DSC rectum +0.3%; LI vs RF: DSC bladder +6.0%). DL performances were confirmed at the DAP comparison (e.g., in CC bladder DAP 3mm : MV 74.5%, ST 27.1%). In both sites, all algorithms showed excellent DSC values for femoral heads contouring ( Figure 1 ). Bowel bag AS showed its dependence on cranial extension. Contouring limitation to the manual cranial extension confirmed DSC increases (e.g. MV DSC BB = 0.67 vs DSC BB_corr = 0.88). An example of CC and PC automatic segmentation of bladder is reported in Figure 2 . In CC, 23.0, 24.0, 10.0, and 12.0 minutes have been the median correcting times to edit ST, RF, MV, and LI, respectively for CC-RO1. On the other hand, 30.0, 30.0, 5.0, and 7.0, minutes have been necessary to correct ST, RF, MV, and LI by CC-RO2. MV has the best CC qualitative evaluation scoring for both CC-ROs and, in general, DL tools outperformed ST and RF, showing small edits requirement. CC-RO1 assigned a good rating in 0, 1, 5, and 4 out of 5 examined SS for ST, RF, MV, and LI, respectively. CC-RO2 assigned a good score in 0, 0, 5, and 4 out of 5 examined SS for ST, RF, MV, and LI, respectively. In PC, median correcting times registered by PC-RO were 16.0, 20.0, 16.0, and 15.0 minutes for ST, RF, MV, and LI, respectively. PC qualitative evaluation confirmed DSC results showing DL’s excellent performance. LI has proven to have the most desirable contours by having a good agreement in 4 out of 5 cases, although ST and RF showed just small differences to be edited. Results: