ESTRO 2024 - Abstract Book

S5028

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

The original image grid had a median dimension of 512 x 512 x 255, while the region of interest (ROI) measured 20 x 21 x 6. Based on our clinical experience we assumed a non-existent risk of BRN below the radiation dose of 30Gy, hence we cropped the original images (i.e. planning CT, dose map, and pre-treatment T1v, and T2w) by considering a bounding box around voxels that fall above the 30Gy threshold. As a result, the median image size was reduced to 105 x 86 x 38 (cropped image grid), coupled with minor alleviation in the class imbalance. A patch-based 2D UNet architecture was employed for the voxel-wise prediction of BRN. For CT and dose maps, since the intensity values represent quantitative measurements that reflect physical properties, a global z-score normalization scheme was adopted by computing the mean and standard deviation of all the cropped images. On the other hand, since MR images are qualitative and indicate relative intensity values, a local z-score normalization scheme was adopted. The input patch size was configured at 128 x 128 to cover the image’s median in-plane resolution. The UNet model accepts 4-channel inputs where each channel represents the 2D patches of dose map, CT, T1v, and T2w images respectively. The model outputs the probability maps associated with the background and BRN. Pytorch [3] library was used for the development and optimization of the model. Figure 2 illustrates a simplified overall summary of the workflow involved in this study.