ESTRO 2024 - Abstract Book

S5115

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Our data showed changing the segmentation volume impacts feature reliability in phantom. This study has demonstrated that there is variability in the reliability of radiomics features across CT-based preclinical scanners. Yet, normalisation steps can be useful to improve the comparability of CBCT and µCT scans e.g. standardisation of imaging energy and pre-processing factors (voxel size). Overall, µCT imaging produced more reliable features than CBCT in mice, with notable superiority in higher-density tissue such as bone, suggesting analysis of CBCT images should be limited to lower-density tissues. Also, the number of reliable features extracted from both CBCT and µCT scans reduced as tissue density increased from lung to bone. The reliability of radiomics outputs was significantly improved by the application of wavelet filtering. Reliable features identified for lung, heart, and bone tissue were directly compared to each other, and tissue density-specific preclinical radiomics signatures were developed. Specifically, the signatures consisted of 133 features for lung, 35 features for heart, and 15 features for bone, which are shared across both CBCT and µCT modalities.

Conclusion:

We present the first cross-centre comparison of reliable radiomics outputs from two preclinical CT scanners. This study demonstrates the importance of standardisation and reinforces the need for multi-centre radiomics studies to produce transferable and meaningful results.

Keywords: Preclinical, CT, Cross-Centre

2361

Digital Poster

Voxel-wise analysis of DWI lung MR images using deep learning to predict PD-L1 expression

Sithin Thulasi Seetha 1,2 , Chandra Bortolotto 1,3 , Chiara Podrecca 4 , Gaia Messana 1 , Alessandra Marrocco 1 , Riccardo Bellazzi 4 , Andrea Riccardo Filippi 1,5 , Lorenzo Preda 1,3 1 University of Pavia, Diagnostic Imaging and Radiotherapy Unit, Department of Clinical, Surgical, Diagnostic, and Pediatric Sciences, Pavia, Italy. 2 National Center for Oncological Hadrontherapy (CNAO), Clinical Department, Pavia, Italy. 3 Fondazione IRCCS Policlinico San Matteo, Radiology Institute, Pavia, Italy. 4 University of Pavia, Department of Electrical, Computer and Biomedical Engineering, Pavia, Italy. 5 Fondazione IRCCS Policlinico San Matteo, Department of Radiation Oncology, Pavia, Italy

Purpose/Objective:

Programmed death-ligand 1 (PD-L1) status is an important biomarker to stratify patients in immunotherapy. The gold standard approach to assess its expression in the malignant tissue is immunohistochemical analysis which requires an invasive biopsy. In this study, we propose a voxelwise deep-learning model to predict PD-L1 expression among non-small cell lung cancer (NSCLC) patients using high b-value diffusion-weighted magnetic resonance (DWI MR) imaging.