ESTRO 2024 - Abstract Book

S3078

Physics - Autosegmentation

ESTRO 2024

these, 75 belong to Institute 3, same center of model’s training, here used as internal data test. The dataset pre processing is the same applied to the training set. 3D CTs images were resampled to uniformly have spacing of 1*1*1.5 mm 3 and size of 512*512*256 voxels and they were converted into Nifti masks. The images were then cropped into 256*256*256 voxels, choosing the first/fourth quadrant (respectively for left/right side target breasts). Quadrants were defined through symmetric criteria starting from the 3D Body ROI center. Moreover, to enhance the difference between glandular and fat tissue, CTs were rescaled from the range -200 – 100 HU to [0, 1]. The trained 3D Unet [7] was applied to the pre-processed images for prediction. From the generated CTV prediction, quantitative volumetric metrics were extracted to assess the predicted versus clinical contours quality and compare between centers: Dice Similarity Coefficients (DSC), Hausdorff Distance (HD), 95th percentile of the Hausdorff Distance (HD_95) and the volumes difference (VD) of CTVs segmented by radiation oncologists and the predicted ones. The Mann Whitney U test was used to see if DSC results of external centers are statistically different from the ones predicted for Institute 3.

Results:

The 3D Unet-model, used for CTV prediction, shows cross-validation average performances of DSCs: 0.90, HD: 19.8mm, HD_95: 9.12mm, highlighting good agreement with clinical contours. The performances obtained on internal test set for Institute 3 are in accordance with cross-validation ones (black lines in Fig. 1). For all the metrics, Institute 4 and 5 show the best agreement with training metrics; Institute 1 and 7 are in good agreement, even if with higher variance; Institute 10 and 9 show worse consistency. Generally, the model tends to predict CTVs larger than the radiation oncologists’ ones, mainly in the lateral and cranio-caudal side. All institutes are in line with the training institute, except Institute 10, predicting slightly smaller contours, and Institute 9, with larger differences.