S2437
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
The sensitivity, specificity, and accuracy were 73%, 64%, and 65% for M3. ISUP GG group alone showed lower performance with an AUC of 0.65 (log-rank p>0.05) and sensitivity, specificity, and accuracy of 64%, 59%, and 59%. On subgroup analysis of intermediate ISUP GG patients (n=46), multimodal model (M3) sustained meaningful discriminative ability (AUC of 0.73 and log-rank p<0.01).
for external validation (n=108) with median follow-up of 5.2 years. A validated automated lesion detection algorithm was used to segment lesions using T2W MRI, ADC map, and high-b value DWI sequences. T2W MRI shape and texture radiomics were extracted with Pyradiomics. Three models were developed using BF as a binary outcome: (M1) a clinical XGBoost model (XGB), (M2) a radiomics XGB, and (M3) a multimodal XGB. Clinical co-variates included PSA, age, and ADT use. Five-fold cross validation was used to select hyperparameters. Accuracy, sensitivity, specificity, and AUC were calculated for each XGB, and the optimal cutoff was selected using Youden’s J Statistic from training predictions. Kaplan Meier curves with log-rank tests were completed using censored deaths (a potential competing risk) and compared to the clinical baseline, ISUP GG. Subgroup analysis was performed for intermediate risk ISUP GG patients. Results: In the external validation set, the multimodal model (M3) exhibited the highest discrimination with an AUC of 0.70 (log-rank p<0.05) versus the clinical model (M1) and radiomics model (M2) with AUCs of 0.57 (log-rank p>0.05) and 0.65 (log-rank p>0.05), respectively.
Conclusion: We present an automated multimodal deep learning radiomics model for predicting BF in RT patients. We demonstrate that this model may add prognostic information to existing biomarkers and maintains accuracy in an external cohort. Leveraging AI to predict BF from baseline MRI could change PCa treatment and staging. Pending further validation, this model may directly inform risk-dependent clinical decision making. References: 1. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin. 2024;74(1):12–49. 2. Daskivich TJ, Wood LN, Skarecky D, Ahlering T, Freedland S. Limitations of the National Comprehensive Cancer Network® (NCCN®) Guidelines for Prediction of
Made with FlippingBook - Share PDF online