Nuklearmedizin 2020; 59(02): 99-100
DOI: 10.1055/s-0040-1708151
Wissenschaftliche Vorträge
Radiomics
© Georg Thieme Verlag KG Stuttgart · New York

Evaluating a machine learning based tool for the detection of pathological hotspots in whole-body PSMA-PET-CT scans

A Erle
1   Universitätsklinikum Bonn, Klinik für Nuklearmedizin, Bonn,, Germany
,
S Moazemi
2   Universitätsklinikum Bonn AND University of Bonn, Klinik für Nuklearmedizin AND Computer Science Department, Bonn,, Germany
,
M Essler
1   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
1   Universitätsklinikum Bonn, Klinik für Nuklearmedizin, Bonn,, Germany
› Institutsangaben
Weitere Informationen

Publikationsverlauf

Publikationsdatum:
08. April 2020 (online)

 

Ziel/Aim The importance of machine learning (ML) in clinical environment increases constantly aiming at the facilitation of image-based diagnosis. Differentiation of pathological from physiological tracer-uptake in PET/CT images is crucial for diagnosis and treatment. The aim of this study was to establish and validate an ML algorithm for computer-aided diagnosis in clinical practice using the example of prostate cancer based on prostate-specific membrane antigen (PSMA) positron emission tomography/computed tomography (PSMA-PET/CT).

Methodik/Methods Retrospective analysis of PSMA-PET/CTs of 72 prostate cancer patients resulted in a total of 2452 hotspots for training after that volumes of interest (VoI) were manually delineated using Interview Fusion by Mediso. The hotspots were then labeled pathological (1629) or physiological (823) as ground truth (GT). A total of 80 radiomics features were calculated for each hotspot. For the ML analyses, ExtraTrees (ET) Classifier was selected based on grid search and according to previous work [1]. Two different training sets were used to assess the performance of the ML classifier on validation data: 1) data from 30 patients from the previous study, and 2) data from the 72 patients. As the validation data set, 331 hotspots (path.:128, phys.: 203) were delineated in 15 other patients with similar clinical factors as the patients in the training cohort.

Ergebnisse/Results ET trained with the data of 30 patients resulted in a sensitivity of 0,98 and a specificity of 0,59 on the validation set. The training data of 72 patients with ET resulted in a sensitivity of 0,98 and a specificity of 0,90.

Schlussfolgerungen/Conclusions The combination of manual and automated diagnosis showed to be able to predict hotspot labels with a high sensitivity and may, therefore, be an important tool to assist in clinical diagnosis.

 
  • Literatur/References:

  • 1 Moazemi S, Khurshid Z, Essler M, Schultz T, Bundschuh RA. , “Automated detection of pathological lesions in PSMA PET/CT scans in prostate cancer patients: Analyzing the relative importance of different groups of features”, Nuklearmed 2019; 58 ( (02) ): 107 , DOI: 10.1055/s-0039-1683476.