ESTRO 2026 - Abstract Book PART II

S2438

Physics - Radiomics, functional and biological imaging, and outcome prediction

ESTRO 2026

adjustment. Model performance was preliminary evaluated on the hold-out test set using a iterative bootstrap approach.

Limited Life Expectancy in Men with Prostate Cancer. J Urol. 2017 Feb;197(2):356–62. Keywords: Multimodal, External Validation, AI

Digital Poster Highlight 2151

Development and Preliminary Evaluation of a Dosiomic Classifier for Early PSA Decay Kinetics after CyberKnife SBRT in Prostate Cancer Rocco Mottareale 1 , Rossella Di Franco 1 , Donato Pezzulla 2 , Savino Cilla 3 , Marcello Serra 1 , Valentina Borzillo 1 , Esmeralda Scipilliti 1 , Dario Franzese 4 , Sisto Perdonà 4 , Francesco Passaro 4 , Alessandro Izzo 4 , Sabrina Rossetti 5 , Sandro Pignata 5 , Gianluca Ametrano 1 , Valentina d'Alesio 1 , Francesca Buonanno 1 , Cecilia Arrichiello 1 , Simona Mercogliano 1 , Vincenzo Ravo 1 1 Radiation Oncology, Istituto Nazionale Tumori - IRCCS Fondazione "G. Pascale", Napoli, Italy. 2 Radiation Oncology Unit, Responsible Research Hospital, Campobasso, Italy. 3 Medical Physics Unit, Gemelli Molise Hospital-Università Cattolica del Sacro Cuore, Campobasso, Italy. 4 Department of Uro-Gynecology, Istituto Nazionale Tumori - IRCCS Fondazione "G. Pascale", Napoli, Italy. 5 Departmental Unit of Clinical and Experimental Uro-Andrologic Oncology, Istituto Nazionale Tumori - IRCCS Fondazione "G. Pascale", Napoli, Italy

Results: Clustering revealed two distinct PSA decay

phenotypes. Cluster 1 showed slow decay rate (3- month: 0.19 ± 0.22; 6-month: 0.14 ± 0.13), whereas Cluster 2 exhibited a faster decay rate (3-month: 0.66 ± 0.22; 6-month: 0.41 ± 0.12). Exponential fits yielded decay constants k = 0.15 ± 0.02 ( τ = 6.45 ± 0.83 months) for Cluster 1and k = 0.53 ± 0.13 ( τ = 1.88 ± 0.44 months) for Cluster 2. The XGBoost classifier performance was evaluated through an iterative bootstrap analysis (250 iterations) on the test set, achieving a mean ROC-AUC of 0.74 ± 0.12 and accuracy of 0.69 ± 0.09. The XGBoost classifier achieved class-specific precision and recall of respectively 0.70 and 0.78 for Cluster 1, and 0.67 and 0.57 for Cluster 2. Top predictive dosiomic features were shape and texture metrics, including prostate elongation and bladder run-length non-uniformity, which significantly differed between clusters (p = 0.0299 and p = 0.0114, respectively).

Purpose/Objective: This study aimed to evaluate the preliminary

performance of a novel dosiomic model designed to predict prostate-specific antigen (PSA) decay kinetics at 3 and 6 months in patients with favourable low- to intermediate-risk prostate cancer treated with Stereotactic Body Radiotherapy (SBRT) using the

Conclusion: Preliminary results indicate that dosiomic features from CyberKnife SBRT dose distributions can discriminate between two early PSA decay phenotypes and show moderate predictive performance in low- to intermediate-risk prostate cancer. Given the limited sample size and parameter variability, validation on larger, independent cohorts is required prior to reliable clinical implementation. Keywords: Dosiomics, Prostate Cancer, CyberKnife SBRT

CyberKnife system. Material/Methods:

A retrospective cohort of 120 patients treated with CyberKnife SBRT (35–36.25 Gy in five fractions) was analyzed. PSA decay rates at 3 and 6 months were calculated as decay rate at month k as, rk = - (1/k) ln (PSAk / PSApre-RT). Unsupervised K-Means clustering, with the optimal number of clusters determined by silhouette analysis (average silhouette = 0.49), stratified patients based on paired 3- and 6-month decay rates. Treatment dose distributions were converted into dose-series, and dosiomic features were extracted from the prostate (CTV/GTV), planning target volume (PTV), bladder, rectum, and penile bulb. After removing highly correlated features (Pearson r > 0.95) and standardization, 35 informative features were selected via LASSO regression. A binary XGBoost classifier was trained on 80 patients (80% training, 20% test) using 10-fold cross-validation, with class imbalance addressed by SMOTE and class-weight

Digital Poster 2223

Noninvasive Prediction of Local Control Using CT Radiomics and Machine Learning After SABR for Colorectal Pulmonary Metastases Donatella Caivano 1 , Federica Palmeri 2 , Antonella Del Gaudio 3 , Donato Pezzulla 4 , Paolo Bonome 4 , Giuseppina Apicella 1 , Sara Gomellini 1 , Daniela Musio 1 ,

Made with FlippingBook - Share PDF online