ESTRO meets Asia 2024 - Abstract Book

S400

RTT – Treatment planning, OAR and target definitions

ESTRO meets Asia 2024

1 Department of Radiation Oncology, Moritz Kaposi General Hospital, Dr. Jozsef Baka Center, Kaposvar, Hungary. 2 Department of Medical Imaging, University of Pécs, Faculty of Health of Sciences, Kaposvar, Hungary. 3 Doctoral School of Health Sciences, University of Pécs, Faculty of Health Sciences, Pecs, Hungary. 4 Medical Physics Department, Institut Jules Bordet, Bruxelles, Belgium. 5 Department of research, Meidcopus Non Profit Ltd, Kaposvar, Hungary. 6 Department of oncoradiology, University of Debrecen, Faculty of Medicine, Debrecen, Hungary

Purpose/Objective:

Our goal was an objective and subjective comparison of the pelvic lymph regions segmented contours by four DLC-based models in prostate tumor patients using three AI softwares (Mirada*, Mvison**,Limbus***).

Material/Methods:

We selected 10 prostate cases randomly whom were treated at our institute. According to the planned CT institute protocol, with a slice thickness of 3 mm, starting from the L1/2 vertebra and caudal to 3 cm from the ischial tuberosity. For the objective comparison of the four segmentations (Mvision: RTOG, Pivotal, Mirada, Limbus), we used volumetric (Dice similarity coefficient, "DSC") and distance-based (Median surface distance "MSD", 95% Hausdorff distance, "HD95%)" indices. Subjective evaluation was classified into major (significant) and minor (insignificant) categories according to the degree of modification. In addition, we recorded the duration of the modification and manual contouring.

Results:

Based on the pairwise comparison, the value of the Dice similarity coefficient showed a similar value: RTOG vs. Pivotal 0.90 (interval 0.88-0.91), Pivotal vs. Mirada 0.84 (0.81-0.91),Mirada vs. Limbus 0.83 (0.79-0.88), Pivotal- Limbus 0.82 (0.79-0.86), RTOG-Limbus 0.81 (0.79-0.83) while RTOG vs.Mirada, it was 0.81 (0.76-0.84). The median surface distance remained below 2.5 mm on average, considering all cases and comparisons: RTOG vs. Pivotal 1.2 mm (0.9-1.5), Pivotal vs. Mirada 1.8 mm (1.1-2.3), RTOG vs. Mirada 2.4 mm (2.0-3.1), Mirada vs. Limbus 2.1 (1.5-2.7), Pivotal-Limbus 2.3 (1.8-3), RTOG vs. Limbus 2.5 (2.3-2.7). Hausdorff distance 95% resulted in a difference of up to nearly 13 mm: RTOG vs. Pivotal 6.9 mm (4.8-9.1), Pivotal vs. Mirada 6.3mm (3.5-8.3), RTOG vs. Mirada 9.2 (6.5-12.3), Mirada vs. Limbus 7.1 (4-8.5), Pivotal vs. Limbus 8.3 (5.5-11.3), RTOG-Limbus 9.7 (8.4 11). In the case of subjective evaluation, RTOG 4-5, Pivotal 3-3 , Limbus 3-3 and Mirada 7-4 required relevant (major) or minor correction overall. The use of the three software significantly shortened the contouring time of the prostate lymph node by an average of 1.5 minutes, compared to manual segmentation, which required an average of 9 minutes. In addition, it was observed that the auto-segmentation usually produced the same deviations compared to our own institute's contouring guidelines, so the correction could be carried out in a more targeted manner, which further reduces the time needed to correct the contour. Mvison can perform segmentation based on two international protocols (RTOG and Pivotal). In our experience, the model following European guidelines required less correction.

Conclusion:

Auto-segmentation (DLC) models based on deep learning significantly reduced the time required for contouring, as they segment the pelvic lymph node region with a good approximation. We found an excellent and good match between the four models, overall the MVision Pivotal and Limbus model requires the least amount of modification.