ESTRO 2023 - Abstract Book

S1885

Digital Posters

ESTRO 2023

Sophia Antipolis, France; 15 IEO European Institute of Oncology IRCCS, Precision Imaging and Research Unit - Department of Medical Imaging and Radiation Sciences, Milan, Italy Purpose or Objective The most common systems used to risk-stratify PCa patients has been repeatedly shown to have suboptimal prognostic and low-end stratification performances, possibly leading to over- or under-treatment of the patients. The aim of the present study is to test the ability of high-performance mathematical models employing clinical radiological and radiomics features to improve the accuracy of non-invasive prediction of pathological features of PCa and therefore improving patients’ stratification in a large cohort of patients treated and managed at the same institution. Materials and Methods A cohort of 949 patients who have undergone mpMRI of the prostate and prostatectomy in IEO between 2015 and 2018 was selected. The prostate gland was segmented with an internally-developed deep learning segmentation algorithm. Gradient- boosted decision tree models were separately trained using pre-treatment clinical features alone and in combination with radiological and/or radiomic features to predict pathological scores, as well as biochemical and clinical progression ( Figure 1 ). Models were validated with 32-times repeated 5-fold cross-validation and evaluated in terms of their AUC values. The behavior of all features-model within different risk groups and PI-RADS categories was analyzed, assessing radiomic contribution through the cumulative SHAP value and the mean absolute error (MAE). A comparison regarding misclassified patients between the models and the clinical workflow was performed as well.

Results The AUC performance of the four models can be seen in Figure 2a . The model including all variables resulted the best model in most endpoints. Radiomics appear to bring a measurable boost in model performances, although small. Considering the model including all features, SHAP subgroup analyses showed that, although the mean/median influence of radiomic features is low, their contribution to individual patients prediction can be very high; moreover, MAE values resulted lower in low-risk and low-PIRADS classes ( Figure 2b ). The best prediction model outperformed the naïve one in all the considered endpoints in terms of AUC, whereas the accuracies were comparable ( Figure 2c ).

Conclusion