S2488
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
features, with the aim of further personalising treatment planning. References: [1] CANTERLA, G.; ACHA, B.; BORREGO, M.; SERRANO, C.; LÓPEZ GUERRA, J.L. Radiomics-Based Characterization of Hematologic Toxicity in Lung- Cancer Radiotherapy. In: 25th International Conference on Bioinformatics and BioEngineering (BIBE 2025). Athens, Greece, 6–8 November 2025.[2] CANTERLA, G.; ACHA, B.; BORREGO, M.; LÓPEZ, J.L.; SERRANO, C. Hacia una predicción personalizada de toxicidad en cáncer de pulmón: integración de radiómica, dosimetría y biomarcadores genéticos. In: XLIII Congreso Anual de la Sociedad Española de Ingeniería Biomédica (CASEIB 2025). Zaragoza, Spain, 19–21 November 2025. Keywords: Radiomics, Bone Marrow, Hematologic Toxicity Limitations of CT-based radiomics when validated externally: Models trained on prostate cancer patients from 2 centres (N = 1298) for BCR prediction Peter D McHale 1 , Fereshteh Gholami 1 , Alan McWilliam 2 , Jane Shortall 2 , Joe M O'Sullivan 1,3 , Suneil Jain 1,3 , Conor K McGarry 1,4 1 Johnston Cancer Research Centre, Queen's University Belfast, Belfast, United Kingdom. 2 Department of Cancer Sciences, The University of Manchester, Manchester, United Kingdom. 3 Clinical Oncology, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom. 4 Radiotherapy Physics, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom Purpose/Objective: CT-based radiomics has shown promise in prediction of prostate cancer risk group [1] but work to date has lacked external validation. This work investigated CT- Digital Poster Highlight 4523 based radiomics for prediction of biochemical recurrence (BCR), and the effectiveness of models when externally validated. Material/Methods: Two datasets of prostate radiotherapy planning CTs (pCTs) were collected from two UK centres. BCR (Phoenix definition [2]) was also recorded for both datasets. Dataset 1 consisted of 352 pCTs for external beam radiotherapy (EBRT) prostate patients diagnosed from 2003-2009. Median slice thickness was 2.5mm (range: 2.5-5mm). Dataset 2 featured 946 patients (775 EBRT, 171 brachytherapy) diagnosed from 2005-2013. Median slice thickness was 5mm (range: 2.5mm-5mm). An additional 20 patients from the same centre as
dimensionality and retain significant features. For each of the nine available HT (grade ≥ 2) endpoints, 2 to 12 features were retained in this way. Subsequently, a multivariate analysis was performed. To overcome the imbalance problem of individual endpoints, the nine endpoints were combined into a single binary outcome (any HT). The input features for this multivariate model were the union of all the features that were significant for each individual endpoint, producing a compact panel of 19 BM radiomic descriptors. No clinical variables were retained as significant. In this model, a class-weighted SVM with an RBF kernel was used as the classifier. To prevent information leakage during the validation process, a 5- fold cross-validation scheme was applied and this process was repeated 5 times. Within each training fold, missing values were handled and the data was standardised. The decision threshold was set to prioritise sensitivity in the event of class imbalance. Hyperparameters were tuned via randomised search (60 iterations) and consolidated through consensus across outer folds. Results: The most informative radiomic features were mainly first-order and zone/dependence-based textures (GLSZM/GLDM), complemented by GLCM, GLRLM, NGTDM, and shape descriptors; notably, variance/non- uniformity, entropy, and high gray-level–emphasis measures predominated. Numerical performance is detailed in Table 1; The model achieved good discrimination and calibration, with an AUC-ROC of 0.741 and a recall of 0.9024 at the pre-specified threshold.
Conclusion: A fully automated BM-radiomics model using a compact 19-feature signature provides predictions of HT in lung-cancer RT, with high sensitivity and acceptable precision. This approach combines validated BM segmentation and radiomics to create a generalisable predictive tool for early risk stratification and workflow prioritisation. Future work will include multicentre external validation, decision-curve analysis and the integration of patient-specific dosimetric
Made with FlippingBook - Share PDF online