ESTRO 2021 Abstract Book

S770

ESTRO 2021

and/or lymph nodes, from 1854 patients were extracted from DICOM RT structure sets and converted to binary masks. An InceptionResNetV2 2D classification network was trained using four different image input channels and a batch size of 72. Input channel 1-3 consisted of binary (i.e. modality independent) orthogonal 2D structure projections from each individual structure, down sampled to 256x256. The fourth channel contained a down sampled summation of every other binary patient structure mask. The network was trained using 10- fold cross validation. Structure classification performance was assessed in CT and MR test datasets with 200 and 40 patients respectively (originating from same clinic as training data), using a majority voting approach including the 10 cross validation models. An external CT dataset with 99 patients from another radiotherapy clinic was also evaluated. Results A class weighted classification accuracy of 99.4% was achieved for training. The unweighted classification accuracy and the weighted average F1 score for different structures in the CT test dataset was 98.8% and 98.4% (n=200, figure 1), and 98.6% and 98.5% (n=40) for the MR test dataset, respectively. For the external CT dataset analyzed for trained structures only, the corresponding results were 98.4% and 98.7% (n=99). The results from the full external CT dataset yielded 79.6% and 75.2%, respectively. Misclassifications in internal CT and MR test datasets were due to existence of multiple CTV structures and structures on which the model was not trained on. Most failures in the external test dataset occurred due to multiple CTVs being grouped into one single PTV, which was not included in the training data.

Conclusion Our proposed method for automated, deep learning-based data cleaning and standardizing of RT prostate OAR and target annotations shows great promise on internal independent CT and MR datasets. Existing structure annotations were automatically renamed according to a user-defined standard and the DICOM header was successfully updated. However, clinic specific contouring methods for the data of interest needs to be represented in the training data for successful application of this method. PD-0929 Volumes and doses by various ways to define rectum in radiotherapy: clinical, standardized, and AI C. Olsson 1 , R. Suresh 1 , S. Akram 2 , J. Niemelä 3 , A. Valdman 4 1 Institute of Clinical Sciences, Sahlgrenska Academy, Gothenburg University, Department of Radiation Physics, Gothenburg, Sweden; 2 MVision AI, Technology, Helsinki, Finland; 3 MVision AI, Products, Helsinki, Finland; 4 Karolinska University Hospital, Department of Radiotherapy, Stockholm, Sweden Purpose or Objective Successful delivery of radiotherapy (RT) relies on correct and consistent definition of organs at risk (OAR). Artificial intelligence (AI) auto-segmentation based on deep learning (DL) is emerging as a means to reduce variability of OARs. Knowledge about how auto-segmentation tools perform on standardized datasets is limited. Our aim is to evaluate the quality of rectal volumes defined by an existing DL auto-contouring algorithm on a standardized dataset for rectal volumes in prostate cancer RT as proposed by the Swedish STRONG guidelines for male pelvis[1] 1 [2] 2 [3] 3 .