S845
Clinical - Lung
ESTRO 2026
Netherlands Cancer Institute, Amsterdam, Netherlands. 4 Computer Science, University of Toronto, Toronto, Canada. 5 Joint Department of Medical Imaging, University Health Network, Toronto, Canada Purpose/Objective: Detection of ILD has significant implications for patients with lung cancer as it is correlated with higher risks of treatment related side effects including radiation pneumonitis.1 A novel algorithm (MIRACLE- ILD) has been developed which can screen patients planned for radiotherapy for ILD using the exhale breathing phase radiotherapy planning computed tomography (CT) scan. Not all radiotherapy centers plan radiotherapy on the exhale breathing phase which may affect MIRACLE-ILD performance in those centers. Material/Methods: Using four dimensional (4D) CT images obtained during treatment planning for curative radiotherapy, the MIRACLE-ILD algorithm was applied to all 4DCT information and the resulting ensemble model scores recorded for each patient. The phases/reconstructions assessed included: ten phases from inhale to exhale, mean/average reconstruction, maximum intensity projection (MIP) reconstruction (used for tumor delineation), and mid- ventilation/position reconstruction. The resulting model scores were summarized for each reconstruction/phase to identify patterns. The performance of the screening test was further assessed by comparing the model outputs for the inhale and exhale 4DCT phases. Statistics were performed to estimate the shift in the scores resulting in coversion of the screening test from negative to positive or vice versa. Results: 162 patients with full 4DCT information were available for analysis. For the MIP reconstruction, the model consistently scored patients higher than the exhale or other 4D phases and was less well correlated (r=0.77) with exhale. Comparing exhale to inhale or mean/average, the ensemble scores remained similar with a high correlation (r = 0.91 and 0.92 respectively). When examining a specific testing threshold comparing inhale and exhale, a proportion of patients who were above threshold on the exhale 4DCT dataset (8/17) had a shift to a lower value below threshold. One patient had the reverse shift (below to above threshold). Overall, fewer patients were model positive (10/162, 6%) on the inhale compared to exhale (17/162, 11%) using the fixed exhale threshold.
Conclusion: The MIRACLE-ILD algorithm performs numerically similarly on most reconstructed phases of the 4DCT but has significant differences in output on the MIP. The 4DCT reconstruction choice affects model results when employing a fixed threshold based testing implementation. This suggests that local calibration to the type of breathing phase of the dataset used for planning will be required before deployment in a new clinical environment. Future AI models trained on specific 4DCT breathing phases may not be extensible to other 4DCT breathing phases without detailed evaluation. References: 1. Bacon, H. et al. Association of artificial intelligence- screened interstitial lung disease with radiation pneumonitis in locally advanced non-small cell lung cancer. Radiother Oncol212, 111144 (2025). Keywords: interstitial lung disease, AI, 4DCT Digital Poster 4538 Re-irradiation with overlapping fields in previously irradiated early-stage lung cancer: Modern techniques, modest adverse events Ipek Sucak, Yunus Babayigit, Oguzhan Bascik, Pantea Bayatfard, Yakup Arslan, Sumerya Duru Birgi, Serap Akyurek Radiation Oncology, Ankara University Faculty of Medicine, Ankara, Turkey
Made with FlippingBook - Share PDF online