ESTRO 2024 - Abstract Book

S3088

Physics - Autosegmentation

ESTRO 2024

First, an image pre-processing was run. All images were re-sampled in order to have uniform images of 512*512*256 voxels (voxel space: 1*1*1.5 mm 3 ). To help in the CTV side identification (right or left) and to reduce the input parameters without reducing the voxel resolution, an automatic cropping of the left/right quadrant was applied, resulting in final images’ dimension of 256*256*256 voxels. To do this, a median of the body ROI center derived in each patient CT slice, applying symmetric criteria, was used. Moreover, to make the network properly working and to enhance the glandular and fat tissue intensity difference, a minimum of -200 HU and maximum of 100 HU were chosen and then scaled to be between 0 and 1. Then, all images (i.e., with left/right BC) were used to train a global cross-validation model. A 3D Unet, enriched with dropout and residual units, from the MONAI library (https://zenodo. org/records/4323059) and implemented in house with python 3.10.9, was used. Zero borders were removed to focus on the valid body area. Data augmentation (random flip on x, rotation, translation, scaling, gaussian sharpening) was performed. The segmentation performance of the cross-validation dataset was addressed through Dice Similarity Coefficient (DSC), Intersection over Union (IOU) and 95th percentile Hausdorff distance (HD_95) metrics. Each 3D segmentation prediction was cropped in the cranio-caudal direction, a posteriori, to match the clinical segmentation, due to their expected large inter-observer variability. Finally, the predicted segmentation was co-registered back to the original image dimension for clinical purpose.

Results:

The best 3D model obtained was able to reach, after 178 epochs, high metrics performance (DSC: 0.903; HD_95: 9.1 mm, see Figure 1), similar to what reached in [9]. Our network, however, is a 3D model and the predictions are smoother between slices, as it should be, well mimicking clinical practice.