ESTRO 2025 - Abstract Book

S2552

Physics - Autosegmentation

ESTRO 2025

3. B., Castonguay, et al. "Stochastic gradient methods with layer-wise adaptive moments for training of deep networks," arXiv preprint arXiv:1905.11286, 2019, doi:10.48550/arXiv.1905.11286.

4207

Digital Poster Deep learning PET/CT-based algorithm for estimating tumor burden in metastatic melanoma patients under immunotherapy Lorenzo Lo Faro 1,2 , Hubert S Gabryś 3 , Simon Burgermeister 3 , Maiwand Ahmadsei 3 , Ciro Franzese 1,2 , Stephanie Tanadini-Lang 3 , Panagiotis Balermpas 3 , Marta Scorsetti 1,2 , Matthias Guckenberger 3 , Sebastian M Christ 3 1 Department of Biomedical Sciences, Humanitas University, Milan, Italy. 2 Department of Radiotherapy and Radiosurgery, Humanitas Research Hospital IRCCS, Milan, Italy. 3 Department of Radiation Oncology, University Hospital and University of Zurich, Zurich, Switzerland Purpose/Objective: Artificial intelligence (AI) is widely used in radiation oncology for tasks like lesion detection and segmentation, but has yet to be integrated for tumor burden (TB) estimation (1–8). Against this background, this study investigates the utility of a PET-based deep learning (DL) software (PARS) to detect and auto-segment metastatic melanoma lesions and estimate global TB accurately. Material/Methods: This retrospective study involved 173 stage IV melanoma patients who underwent PET/CT scans before starting immunotherapy and at 3 and 6 months into follow-up. Gross tumor volumes (GTV) of all metastatic lesions were delineated using both the PARS software with manual segmentations by radiation oncologists as ground truth. Lesions classified as “benign” or “suspicious”. Data collected included lesion number, volume, maximum standardized uptake value (SUV), location, and global TB. Segmentation was performed at baseline (within 90 days of immunotherapy initiation, t.0) and at follow-up time points (3 months, t.1, and 6 months, t.2). Results: The PARS model achieved a 70.3% recall, identifying approximately 70% of all true lesions, with a precision of only 49.1%, indicating a high number of false positives. Precision and recall varied by anatomical site, with lung lesions showing the highest precision (75.4%) and bone lesions the lowest (33.6%). Higher probability thresholds improved precision but reduced recall, with lung lesions achieving the best trade-off. Bone lesions, despite high recall, had low precision, leading to more false positives in that region ( Figure 1 ). For TB estimation, PARS exhibited modest agreement with expert delineations (Pearson’s r=0.31). Liver lesions showed the strongest correlation (r=0.58), while lymph nodes showed the weakest (r=0.28). Discrepancies were notable at lower TB values, where PARS tended to overestimate TB by an average of 97% across time points ( Figure 2 ). Mean absolute relative percent difference (MARPD) values were high for bone (152%) and liver (124%) lesions, indicating substantial relative error, particularly in these regions. Bone lesions showed the smallest mean absolute error (MAE, 9.5 cc), lung and lymph node lesions exhibited higher errors, reflecting greater estimation challenges.