ESTRO 2021 Abstract Book

S1221

ESTRO 2021

Conclusion Due to strong tracer uptake and sharp contrasts in terms of high TBRs in primary and metastatic lesions, 68 Ga- FAPI-PET/CT emerges as an enhancement to standard diagnostic imaging with regard to (radio-)oncological management. However, prospective studies are required to confirm the high potential of the new hybrid imaging probe for rare diseases. PO-1491 A comparison of three commercial AI contouring solutions P. Doolan 1 , S. Charalambous 2 , Y. Roussakis 1 , A. Leczynski 2 , K. Ferentinos 2 , I. Strouthos 2 , E. Karagiannis 2 1 German Oncology Center, Department of Medical Physics, Limassol, Cyprus; 2 German Oncology Center, Department of Radiation Oncology, Limassol, Cyprus Purpose or Objective To compare the performance of three commercial artificial intelligence (AI) contouring solutions. To determine if they offer an advantage compared to manually-drawn expert contours, both in terms of quality and time, and to observe if use of such contours translates to an impact on the dosimetry of the plan. Materials and Methods In this work three commercial AI contouring solutions were tested: DLCExpert from Mirada Medical (Mir) (Oxford, UK); MVision (MVis) (Helsinki, Finland); and Annotate from Therapanacea (Ther) (Paris, France). A total of 80 patients were assessed: 20 from each of the four anatomical sites (breast, head and neck, lung, and prostate). The AI-generated contours were compared to expert contours drawn by Radiation Oncologists following the RTOG and MacDonald et al (2016) protocols. Differences in the structures drawn by the commercial systems were identified. The times to generate the expert contours, as well as the time to correct the AI-generated contours, were recorded. Each AI contour was quantitatively compared to its corresponding manually-drawn expert contour using the DICE index, concordance index and Hausdorff distance. The AI contours from each software were overlaid on the planned radiation dose to observe if the different contours result in any dosimetric differences. Results For the expert contours, the mean time required to manually draw the twenty cases for each site was: 22 mins (breast); 112 mins (head and neck); 26 mins (lung); and 42 mins (prostate). The time taken to correct the AI- generated structures was less than these durations and very similar among the three commercial solutions, for all sites. The number of structures contoured by the manually-drawn Expert and Mir/MVis/Ther solutions was: 10 and 9/17/30 for breast; 33 and 29/46/46 for head and neck; 7 and 5/13/8 for lung; 12 and 7/16/18 for prostate. Compared to the structures that are routinely contoured in our clinic, there were a number of structures missing from each model. Conversely, in the prostate model MVis contours the lymph nodes, while Ther contours the lymph nodes for breast, head and neck and lung models (none of which are not routinely contoured in our clinic). The DICE and concordance indices were very high for all systems, across all sites. Figure 1 shows the DICE indices for breast, which had median DICE values for Mir/MVis/Ther: Heart 0.91/0.96/0.92, Oesophagus - /0.79/0.82, Lung_L 0.98/0.96/0.97, Lung_R 0.97/0.95/0.96, Breast_L 0.88/0.88/0.90, Breast_R 0.82/0.72/0.89, Liver 0.76/0.82/0.84, SpinalCanal -/0.96/0.96. Figure 2 shows the corresponding Hausdorff