Evaluation of AI based contouring tools in prostate cancer RT
PH-0484
Abstract
Evaluation of AI based contouring tools in prostate cancer RT
Authors: Borkvel|, Anni(1)*[Anni.Borkvel@regionaalhaigla.ee];Gershkevitsh|, Eduard(1);Adamson|, Merve(1);Kolk|, Kati(1);Zolotuhhin|, Daniil(1);
(1)North Estonia Medical Centre, Radiotherapy Centre, Tallinn, Estonia;
Show Affiliations
Hide Affiliations
Purpose or Objective
OAR and target contouring is a labour intensive step in an RT treatment planning process. Recently multiple vendors have introduced a segmentation algorithms based on artificial intelligence (AI) to reduce the amount of manual work. In this study, the accuracy and efficiency gain for prostate cancer patient contouring when using AI based contouring tools from two vendors was assessed.
Material and Methods
Contouring accuracy was evaluated for 37 patients by comparing manually contoured structures to automatically segmented structures using MVison contouring tool. Standard Imaging StructSure software was used to evaluate the difference in structure volumes and Dice similarity coefficient. The following structures were segmented and evaluated for all patients: bladder, rectum, prostate, seminal vesicles, femoral heads and penile bulb.
To assess the efficiency gain, two experienced RTTs produced manual contours for 18 patients and recorded the time required for outlining. Different set of 15 patients was automatically segmented, then manually edited by the same RTTs and time required recorded. Use of TPS semi-automatic segmentation tools was allowed. Average time per patient was calculated for manual contouring and editing task.
To compare 2 different vendors the same 4 prostate patients were segmented using MVision and Mirada Medical contouring tools.
Results
Table 1 shows results of the automatic segmentation. Difference in volume for bladder and rectum is less than 10% and Dice coefficient is higher than 0.8. For the rest of the structures the difference in volumes is greater (27.0%-69.5%) and reflected in a poorer Dice coefficient results (0.44-0.78). For prostate, MVision contouring tool systematically added extra contour on one CT slice above and below those outlined manually. The larger difference for femoral heads are due to the different definition of these structures in the AI training set.
![](https://cmoffice.estronet.eu/cmAbstractSubmission/Content/cmabstsms/E39/abstracts/0000005400/0000632300/Images/Table600x653PKyleval.png)
Manual contouring took on average 54.9 min, editing automatically segmented structures required on average 43.1 min per patient. Therefore, the use of MVision automatic segmentation algorithm reduced the contouring time by 21.5%.
When comparing manual contouring time vs MVison produced contour editing vs Mirada Medical produced contour editing for 4 prostate cancer patients, then MVison was requiring 28.1% and Mirada Medical 4.9% less editing time than manual contouring. Dice scores were comparable between MVison and Mirada Medical contouring tools.
Conclusion
While the manual editing is still needed, AI based contouring reduces the time required. It should be noted that efficiency gain, among other factors, depends on RTTĀ“s experience and therefore, it can be assumed that involving RTTs with less experience would give a better time efficiency gain. The structure definition in the training set is important and could be used to harmonise contouring practices among different clinics.