Nuklearmedizin 2020; 59(02): 170-171
DOI: 10.1055/s-0040-1708365
Wissenschaftliche Poster
Radiomics
© Georg Thieme Verlag KG Stuttgart · New York

Analyzing different combinations of radiomics features and clinical data for treatment response prediction based on whole-body PSMA-PET-CT scans: A machine learning based approach

S Moazemi
1   Universitätsklinikum Bonn AND University of Bonn, Klinik für Nuklearmedizin AND Computer Science Department, Bonn, Germany
,
A Erle
2   Universitätsklinikum Bonn, Klinik für Nuklearmedizin, Bonn, Germany
,
M Essler
2   Universitätsklinikum Bonn, Klinik für Nuklearmedizin, Bonn, Germany
,
T Schultz
3   University of Bonn, Computer Science Department AND Bonn-Aachen International Center for Information Technology (B-IT), Bonn, Germany
,
RA Bundschuh
2   Universitätsklinikum Bonn, Klinik für Nuklearmedizin, Bonn, Germany
› Author Affiliations
Further Information

Publication History

Publication Date:
08 April 2020 (online)

 

Ziel/Aim Machine learning (ML) has gained critical importance in diagnosis and therapy planning in recent years. In addition to conventional clinical data, radiomics features extracted from prostate-specific membrane antigen (PSMA) positron emission tomography/computed tomography (PSMA-PET/CT) scans can facilitate diagnosis and treatment of prostate disease. Hence, the true screening of patients for available treatment methods based on ML methods is essential. This study investigates the relative impact of different parameters to predict responders to 77Lu-PSMA treatment using ML methods. To this end, the difference in prostate-specific antigen (PSA) levels at pre- and post-therapy (delta PSA) was used.

Methodik/Methods A cohort of 72 prostate cancer patients (59 responders) was analyzed retrospectively. Per-patient average values of 80 radiomics features as well as 22 clinical parameters were analyzed. Linear and non-linear regression, as well as classification based on four ML algorithms (support vector machine (SVM) with linear, polynomial and radial basis function (RBF) kernels and ExtraTrees (ET)) applied to correlate with delta PSA as the therapy response indicator. Grid search and stratified k-fold cross-validation (CV) applied.

Ergebnisse/Results Parameters from both radiomics (e.g., Size Variation) and clinical (Radiotherapy of LN) subsets strongly correlate with delta PSA (p_values<0.05).

ET showed best performance on the combination of 80 PET/CT radiomics features and 22 clinical parameters (81% mean accuracy (ACC) for PSA classification and 0.98 R2 score of regression test). The ACC could be maintained on a balanced subset with 28 subjects (15 responders).

Schlussfolgerungen/Conclusions Machine learning holds promise for patient selection for 77Lu-PSMA treatment, considering combinations of radiomics features and clinical data.