Using a Lymphoma and Lung Cancer Trained Neural Network to Predict the Outcome for Breast Cancer on FDG PET/CT Data

D Kersting; M Weber; L Umutlu; M Schäfers; C Rischpler; WP Fendler; I Buvat; K Herrmann; R Seifert

doi:10.1055/s-0041-1726789

Nuklearmedizin - NuclearMedicine, Table of Contents

Nuklearmedizin 2021; 60(02): 161
DOI: 10.1055/s-0041-1726789

WIS-Vortrag

Radiomics

Using a Lymphoma and Lung Cancer Trained Neural Network to Predict the Outcome for Breast Cancer on FDG PET/CT Data

D Kersting

¹University Hospital Essen, Department of Nuclear Medicine, Essen

,

M Weber

¹University Hospital Essen, Department of Nuclear Medicine, Essen

,

L Umutlu

²University Hospital Essen, Department of Diagnostic and Interventional Radiology and Neuroradiology, Essen

,

M Schäfers

³University Hospital Munster, Department of Nuclear Medicine, Munster

,

C Rischpler

¹University Hospital Essen, Department of Nuclear Medicine, Essen

,

WP Fendler

¹University Hospital Essen, Department of Nuclear Medicine, Essen

,

I Buvat

⁴Institut Curie, Université Paris Saclay, Laboratoire d’Imagerie Translationnelle en Oncologie, Inserm, Paris

,

K Herrmann

¹University Hospital Essen, Department of Nuclear Medicine, Essen

,

R Seifert

¹University Hospital Essen, Department of Nuclear Medicine, Essen

› Author Affiliations

Abstract

Full Text

Ziel/Aim Automatic quantification of the metabolic tumor volume (MTV) from PET data by neural networks can simplify the clinical implementation of the parameter, whose manual estimation is a laborious process. It remains unclear how specific a neural network has to be designed for the investigated disease. The aim of this study was to evaluate the accuracy of a neural network that was trained on lymphoma and lung cancer FDG PET/CT data in the evaluation of breast cancer patients.

Methodik/Methods F-18-FDG PET/CT data sets of 50 breast cancer patients were included. The PARS neural network (Siemens) was used to detect pathological foci and determine their anatomical location. The PARS network was developed on lymphoma and lung cancer PET data. For diagnostic reference consensus reads of two nuclear medicine physicians and follow up data were used. In total 1072 foci were manually segmented. The accuracy was calculated for lesion detection, MTV quantification, and anatomical position determination.

Ergebnisse/Results For PERCIST measurable foci, the per patient sensitivity and specificity of the neural network for lesion detection was high (92 %,CI = 79-97 % and 98 %,CI = 94-99 %). Extending the analysis to all FDG-avid foci, the sensitivity decreased (39 %,CI = 30-50 %). A high correlation was observed between the AI derived and the manually segmented MTV (R²=0.91;p < 0.001). The anatomical position determination was accurate on the body part (98 %;CI = 95-99 %), region (88 %;CI = 84-90 %) and subregion level (79 %;CI = 74-84 %). Moreover, the AI-derived whole-body MTV was a significant prognosticator for overall survival (HR = 1.275;CI = 1.208-1.713;p < 0.001). In a multivariate analysis, also the AI-derived lymph node MTV (HR = 1.190;CI = 1.022-1.384; p = 0.025) and liver MTV (HR = 1.149; CI = 1.001-1.318; p = 0.048) were predictors for overall survival.

Schlussfolgerungen/Conclusions The neural network that was trained on lymphoma and lung cancer data displayed a high sensitivity and specificity for the detection of PERCIST measurable lesions and a high accuracy in anatomical position determination in FDG PET/CT data from breast cancer patients. Nevertheless, the whole body MTV was a significant predictor for overall survival.