ICHNO-ECHNO 2022 - Abstract Book

S31

ICHNO-ECHNO 2022

Conclusion Certain deformable registration software packages, such as those provided by ADMIRE, may be favorable for registering inter-sequence T2 weighted and DWI images. These results are important to ensure the appropriate selection of post- acquisition registration strategies for MR-guided radiotherapy applications seeking to utilize anatomical and functional sequences.

PO-0063 External validation of a 3D ResUNet model to segment oropharyngeal tumors using PET-CT images

K. Wahid 1 , S. Ahmed 2 , A. Mohamed 1 , C. Fuller 2 , M. Naser 1

1 MD Anderson Cancer Center, Radiation Oncology, Houston, USA; 2 MD Anderson Cancer Center, Radiation Oncology , Houston, USA Purpose or Objective Accurate primary gross tumor volume (GTVp) auto-segmentation is a crucial unmet need in radiation oncology workflows for oropharyngeal cancer (OPC) patients. Deep learning (DL) approaches show promise for auto-segmentation of OPC GTVps in PET/CT in part due to the availability of large datasets from the MICCAI HECKTOR Challenge. However, there is a need to validate these methodologies on additional external datasets from independent institutions to ensure adequate generalizability of DL models and facilitate their clinical implementation. Materials and Methods We used pre-treatment PET/CT scans for 325 OPC patients (224 train, 101 test) from the 2021 HECKTOR Challenge, which originated from multiple European and Canadian institutions with GTVp masks generated by multiple experts. In addition, we utilized an independent PET/CT test dataset of 68 OPC patients from MD Anderson Cancer Center (MDA) with GTVp masks generated by a single expert and verified by an additional expert. Images were cropped with a fixed size bounding box of 144 mm 3 , resampled, and normalized. DL models were trained using the 224 training cases from the HECKTOR challenge and separately evaluated on the 101 HECKTOR and 68 MDA test sets. A DL model based on the 3D ResUnet architecture was implemented in the MONAI software package. We evaluated two variants of the model: 256 channels or 512 channels in the bottleneck layer. We implemented data augmentation (flips, rotations) to mitigate overfitting. We used an Adam optimizer with a Dice similarity coefficient (DSC) loss function. For model training, we used a 10-fold cross- validation approach. For each test set, we implemented two model ensembling approaches applied to cross-validation training data to estimate the final tumor masks: Simultaneous Truth and Performance Level Estimation (STAPLE) and majority agreement (AVERAGE). Mean DSC and median 95% Hausdorff distance (95% HD) were used to quantify model performance on the test sets. Results GTV segmentation performance was measured across both the HECKTOR and MDA test datasets ( Table 1 ). On the HECKTOR test set, the 256 AVERAGE method achieved the best overall results with a mean DSC of 0.770 and median 95% HD of 3.143 mm. On the MDA test set, the 512 STAPLE method achieved the best mean DSC result of 0.744 while the 256 STAPLE and 256 AVERAGE methods both achieved the best median 95% HD result of 5.094 mm.