EJNMMI Research (Aug 2021)

Preselection of robust radiomic features does not improve outcome modelling in non-small cell lung cancer based on clinical routine FDG-PET imaging

  • Carol Oliveira,
  • Florian Amstutz,
  • Diem Vuong,
  • Marta Bogowicz,
  • Martin Hüllner,
  • Robert Foerster,
  • Lucas Basler,
  • Christina Schröder,
  • Eric I. Eboulet,
  • Miklos Pless,
  • Sandra Thierstein,
  • Solange Peters,
  • Sven Hillinger,
  • Stephanie Tanadini-Lang,
  • Matthias Guckenberger

DOI
https://doi.org/10.1186/s13550-021-00809-3
Journal volume & issue
Vol. 11, no. 1
pp. 1 – 12

Abstract

Read online

Abstract Background Radiomics is a promising tool for identifying imaging-based biomarkers. Radiomics-based models are often trained on single-institution datasets; however, multi-centre imaging datasets are preferred for external generalizability owing to the influence of inter-institutional scanning differences and acquisition settings. The study aim was to determine the value of preselection of robust radiomic features in routine clinical positron emission tomography (PET) images to predict clinical outcomes in locally advanced non-small cell lung cancer (NSCLC). Methods A total of 1404 primary tumour radiomic features were extracted from pre-treatment [18F]fluorodeoxyglucose (FDG)-PET scans of stage IIIA/N2 or IIIB NSCLC patients using a training cohort (n = 79; prospective Swiss multi-centre randomized phase III trial SAKK 16/00; 16 centres) and an internal validation cohort (n = 31; single centre). Robustness studies investigating delineation variation, attenuation correction and motion were performed (intraclass correlation coefficient threshold > 0.9). Two 12-/24-month event-free survival (EFS) and overall survival (OS) logistic regression models were trained using standardized imaging: (1) with robust features alone and (2) with all available features. Models were then validated using fivefold cross-validation, and validation on a separate single-centre dataset. Model performance was assessed using area under the receiver operating characteristic curve (AUC). Results Robustness studies identified 179 stable features (13%), with 25% stable features for 3D versus 4D acquisition, 31% for attenuation correction and 78% for delineation. Univariable analysis found no significant robust features predicting 12-/24-month EFS and 12-month OS (p value > 0.076). Prognostic models without robust preselection performed well for 12-month EFS in training (AUC = 0.73) and validation (AUC = 0.74). Patient stratification into two risk groups based on 12-month EFS was significant for training (p value = 0.02) and validation cohorts (p value = 0.03). Conclusions A PET-based radiomics model using a standardized, multi-centre dataset to predict EFS in locally advanced NSCLC was successfully established and validated with good performance. Prediction models with robust feature preselection were unsuccessful, indicating the need for a standardized imaging protocol.

Keywords