Nuklearmedizin 2019; 58(02): 107
DOI: 10.1055/s-0039-1683476
Wissenschaftliches Programm: Leuchtturm-Sitzungen
Leuchtturm-Sitzung 2: Radiomics
Georg Thieme Verlag KG Stuttgart · New York

Automated detection of pathological lesions in PSMA PET/CT scans in prostate cancer patients: analyzing the relative importance of different groups of features

S Moazemi
1   Unversitätsklinikum Bonn AND Unversity of Bonn, Klinik und Poliklinik für Nuklearmedizin AND Depertment of Computer Science, Bonn
,
Z Khurshid
2   Nuclear Medicine, Oncology and Radiotherapy Institute, Department of Nuclear Medicine, Islamabad, Pakistan
,
M Essler
3   Universitätsklinikum Bonn, Klinik und Poliklinik für Nuklearmedizin, Bonn
,
T Schultz
4   University of Bonn, Bonn Aachen International Center for Information Technology (BIT) AND Department of Computer Science, Bonn
,
RA Bundschuh
3   Universitätsklinikum Bonn, Klinik und Poliklinik für Nuklearmedizin, Bonn
› Author Affiliations
Further Information

Publication History

Publication Date:
27 March 2019 (online)

 

Ziel/Aim:

Prostate cancer is a common cause of males' death worldwide and prostate-specific membrane antigen (PSMA) images with positron emission tomography/computed tomography (PET/CT) are exclusive bio-markers for this disease. Distinguishing pathological from physiological uptakes in PET/CT images is therefore essential for diagnosis and treatment. Thus, achieving higher accuracy in classifying lesions is critical. The goal of this study was to investigate the role of different feature groups of PSMA PET/CT in classifying hotspots in prostate cancer patients undergoing base-line 68Ga-PSMA followed by 177Lu-PSMA therapy.

Methodik/Methods:

Retrospective scans from 30 prostate cancer patients resulted in 958 hotspots localized using InterView Fusion (Mediso Medical Imaging). For each lesion, 80 features were extracted and categorized into 6 groups: first or higher order statistics, shape/size based, textural, zone length based, run length based and combined. Each group had PET, CT and PET/CT subgroups. For the classification, four different machine learning (ML) classifiers (linear and polynomial kernel support vector machine (SVM), random forest (RF) and extra trees (ET)) were applied. Grid search, leave one subject out cross validation and permutation test were performed to ensure the significance of the results.

Ergebnisse/Results:

We reached a high area under the curve (AUC > 85%) in all the cases except for the shape/size based group (63%). For most of the feature groups, the PET/CT was the best subgroup. The textural features group including heterogeneity parameters was the best with ET (96%). This even outperformed all 80 features used with the same method (95%). The p-value of the permutation test was 0.00125.

Schlussfolgerungen/Conclusions:

Four ML methods applied to rank different feature groups for classifying pathological uptakes in PSMA PET/CT scans. We showed that combining PET/CT features classifies the hotspots with high accuracy. Moreover, the importance of textural features in detecting malignant tissues was shown.