S2300
Physics - Machine learning and AI algorithms
ESTRO 2026
Results:
https://doi.org/10.3310/hta20400.Mikhael, P. G. et. al. (2023). ‘Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography’, Journal of Clinical Oncology, 41 (12), pp. 2191-2200World Health Organisation. (2023). Lung Cancer, 26 June 2023, World health Organisation. Available at: https: www.who.int/news-room/factsheets/detail/lung- cancer Keywords: Lung Cancer, Prediction, Machine learning AI-based metrics of contouring consistency in breast cancer patients predict Radiotherapy outcome Gabriele Palazzo 1 , Maria Giulia Ubeira-Gabellini 1 , Antonella del Vecchio 1 , Andrei Fodor 2 , Claudio Fiorino 1 1 Medical Physics, IRCCS San Raffaele Scientific Institute, Milan, Italy. 2 Radiation Oncology, IRCCS San Raffaele Scientific Institute, Milan, Italy Digital Poster 2983 Purpose/Objective: In breast radiotherapy, variability in the delineation of Clinical Target Volumes (CTVs) can reflect patient anatomy, individual and institutional practice in manual delineation and adherence to giudelines. The increasing availability of robust deep-learning contouring models may enable objective analyses of contours.This study aimed to evaluate whether inconsistency between clinical and AI-predicted CTVs can act as an imaging-derived biomarker for post- radiotherapy outcome, focusing on oedema within 6 months from Radiotherapy and 3-year local relapses. Material/Methods: CT scans and clinical data from 1491 breast cancer patients (2009-2022)[1,2] treated with whole-breast radiotherapy (40Gy/15fr) were analyzed, excluding patients with silicone implants. Local relapse and oedema information was available for 1196 and all 1491 patients respectively. CTVs were predicted using an in-house 3D-UNet model trained on a subset of the cohort used in this study[3] and compared to clinical CTVs via Dice Similarity Coefficient (DSC), total volume difference (VD), and cranial/caudal differences (VDcra/VDcau), with a positive sign if the predicted volumes were more extended than treatment ones. Metrics were computed using a MONAI-based python script[4].The dataset was split in train-test populations with an 80:20 ratio for oedema (476/1491 events). Only training was performed for local relapse due to the low number of events (13/1196). Multiple classifiers (Logistic Regression, XGBoost, HistogramGradientBoost, LightGBM)[5,6] were optimized with 5-fold cross-validation on the training set with Optuna[7], using mean validation AUC for
Figure 1 shows discriminative abilities of all three models with probabilistic binary classification. The weighted average ensemble approach has shown highest AUC followed by clinical risk-based model and image-based model.
Figure 2 shows each model’s ability to accurately predict whether a participant would be diagnosed with lung cancer in next three years’ time. The weighted average ensemble approach achieved an optimal reduction in falsely reported cases. Conclusion: This study demonstrated that the prediction of lung cancer incidence three years in advance is possible using routinely collected sequential LDCT scans and clinical risk features, even when little or no visible cancer nodule information is available at the time of lung screening. References: Field, J. K. et. al. (2016). ‘The UK Lung Cancer Screening Trial: a pilot randomised controlled trial of low-dose computed tomography screening for the early detection of lung cancer’, Health technology assessment (Winchester, England), 20 (40), p. 1.
Made with FlippingBook - Share PDF online