S2486
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
and NL receiving >30/50/60 Gy (NL>30Gy/NL>50Gy/NL>60Gy: n=7/1/3). Model performance decreased when applied to the external cohort: The median (range) AUC was 0.84 (0.72- 0.89) vs. 0.56 (0.49-0.59) for RP and 0.57 (0.45-0.66) for RPEarly. The largest Δ AUC (AUCpublished − AUCexternal) was observed for NL30Gy (RP; Δ AUC=0.40), while the smallest Δ AUC (RP, RPEarly) was seen in the dose-map model ( Δ AUC=0.16, 0.06). Conclusion: The four published RP radiomics signatures for cCRT did not generalize for RP or RPEarly after cCRT and IO in the independent cohort: performance decreased to a median of 67%. While the DWL-based model provided the smallest performance decline, these results highlight the need for new, potentially dose- stratified signatures, enabling using pretreatment CT- based radiomics for routine baseline RP and RPEarly risk prediction after cCRT and IO. Keywords: radiomics, pneumonitis, immunotherapy Digital Poster Highlight 4438 Multi-sequence MRI-radiomics for non-invasive differentiation of tumour progression and radionecrosis in radiosurgery patients Michael Maddalena 1,2 , Simon Blasby 2 , Paula Alcaïde- Leon 3,4 , David Shultz 2,5 , Catherine Coolens 1,2 1 Medical Biophysics, University of Toronto, Toronto, Canada. 2 Department of Radiation Oncology, University of Toronto, Toronto, Canada. 3 Medical Imaging, University of Toronto, Toronto, Canada. 4 Joint Department of Medical Imaging, Toronto Western Hospital, Toronto, Canada. 5 Institute of Medical Science, University of Toronto, Toronto, Canada Purpose/Objective: Despite excellent local control of brain metastases (BM) treated with stereotactic radiosurgery, potential tumour progression (TP) is observed in up to 20% of cases during follow-up1. Radiosurgery treatment may also induce radionecrosis (RN), a radiosurgery by- product indistinguishable from TP via MRI yet requiring distinct management approaches2. Currently, invasive post-surgical histopathology remains the only gold-standard confirmation of TP/RN due to the historical subpar performance of image- based protocols3. Robust non-invasive imaging techniques are urgently required to improve case stratification while minimizing harm. To address this need, we employ an AI-based classification framework leveraging various imaging biomarkers acquired via a multi-sequence MRI protocol, incorporating BM characteristics reflecting lesion perfusion, cellularity and permeability.
modeling was repeated using study's specified procedures for RP and RPEarly. Results:
Four studies published in 2021-2024 (N=60-236) were included. Their radiomics included shape/texture/first- order (n=10/11/1) features from CTs, texture features from radiation dose distributions (n=1), and texture/first-order (n=9/6) features from wavelet- filtered CTs of the normal lung (NL; n=12), whole lung (WL; n=4), WL radiation dose distributions (DWL; n=1), WL receiving 5/20 Gy (WL5Gy/WL20Gy: n=1/1), WL receiving >20/>30 Gy (WL>20Gy/WL>30Gy: n=7/1),
Made with FlippingBook - Share PDF online