ESTRO 2024 - Abstract Book

S5116

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Material/Methods:

We retrospectively collected DWI scans of 31 patients with pathologically confirmed stage III NSCLC acquired with different b-values over the period from April 2021 until February 2023. During radiological assessment, DWI scans obtained at the highest b-value (b=800 s/mm2) exhibited a clear distinction between the tumor and its boundaries, hence were utilized for subsequent analysis. The MR visible lesions were segmented by a skilled radiologist; the PD L1 expression of the tumor was assessed using immunohistochemical analysis which served as the ground truth label along with the tumor annotation. Patients were categorized into two groups: PD-L1 negative (PD-L1 expression <1%, n=14) and positive (≥1%, n=17), thresholded based on the routine clinical stratification of patients to be eligible for immunotherapy. The patient characteristics and imaging acquisition protocols are summarized in Figure 1.

The image grid had a median dimension of 160 x 160 x 39, while the region of interest (ROI) had dimensions of 33 x 29 x 10. To predict PD-L1 expression at the voxel level, a patch-based 2D UNet architecture was employed. Since MR images represent qualitative information with relative intensity values, a local z-score normalization scheme was used as a preprocessing step before feeding the images to the model. The input patch size was configured at 32 x 32 to cover the median in-plane resolution of the ROI. The model predicts the probability of each voxel belonging to PD-L1 negative and positive classes. Pytorch [1] library was used for the development and optimization of the model.