ESTRO 2024 - Abstract Book

S3130

Physics - Autosegmentation

ESTRO 2024

(multi label) on planning PET/CT images. The network was trained on 301 OPC patients treated with (chemo)radiation from 2014 to 2022 in our institute. Internal validation was performed using 5-fold cross-validation. Segmentation performance was assessed on 2 test cohorts: 104 OPC patients treated between 2010 and 2013 at the same institute, and 519 OPC patients from the publicly available 2022 HECKTOR challenge dataset collected from 7 different centres. For each patient in the test sets, a 3D tumour probability map was generated for both GTVp and GTVln. Multiple tumour contours were created for nine different probability threshold values per patient, which were quantitatively compared with manual GTVp and GTVln segmentations by the radiation oncologist using the surface dice similarity coefficients (surface-DSC). The optimal probability thresholds ( optimal threshold ) for the GTVp and the GTVln for each patient were defined as those resulting in the highest surface-DSC. To demonstrate how the optimal threshold values per patient relate to tumour and PET imaging characteristics, patients were grouped based on their optimal threshold values, and the mean values of the GTVp volume in mm 3 , median SUV within the GTVln, and minimum SUV within GTVp were calculated for each optimal threshold group.

Results:

Applying the same probability thresholds to all patients, the mean surface-DSC scores in the internal and external test sets were 0.72±0.23 and 0.75±0.25 for GTVp, and 0.60±0.33 and 0.69±0.30 for GTVln. Conversely, when simulating the intended use of adaptive thresholding by selecting the optimal probability threshold for each patient in both the internal and external sets, the average surface-DSC scores improved to 0.80±0.17 and 0.86±0.22 for GTVp, and 0.65±0.33 and 0.73±0.31 for GTVln segmentation. In Figure 1 , some qualitative examples are presented, showcasing scenarios where distinct optimal thresholds were selected. These individualized thresholds can serve as valuable initial guidance for radiation oncologists in the segmentation process for each patient.