[OA047] Robust radiomic feature selection in magnetic resonance apparent diffusion coefficient maps of rectal cancer. (August 2018)

• Record Type: Journal Article
• Authors: Wee, Leonard; Traverso, Alberto; Kazmierski, Michal; Kalendralis, Petros; Nissen, Henrik Dahl; Dekker, Andre
• Is Part Of: Physica medica. Volume 52(2018)Supplement 1
• Page Start: 20
• Publication Date: 2018-08
• Subjects: Medical physics -- Periodicals; Biophysics -- Periodicals; Biophysics -- Periodicals; Imagerie médicale -- Périodiques; Radiothérapie -- Périodiques; Rayons X -- Sécurité -- Mesures -- Périodiques; Physique -- Périodiques; Médecine -- Périodiques; 610.153
• DOI: 10.1016/j.ejmp.2018.06.119
• Languages: English
• ISSNs: 1120-1797

Abstract

Abstract : Purpose: Changes in Apparent Diffusion Coefficient (ADC) intensities derived from Magnetic Resonance (MR) imaging have been shown to be potentially predictive of (chemo) radiotherapy response in rectal cancer. Radiomics is a non-invasive, computer-assisted extraction of image-based biomarkers that assist in characterization of tumour phenotype. This study concerns the identification and selection of robust (i.e. repeatable and reproducible) radiomic features in MR-ADC across independent scanners in The Netherlands and Denmark. Methods: Retrospective datasets (23 cases from the THUNDER clinical trial and 34 cases from a Danish watchful-waiting protocol) were independently manually delineated. Radiomic features were extracted using public open source software. Repeatability was examined by test-retest using scans 5 min apart. Reproducibility was assessed using concordance and intra-class correlation coefficients for sensitivity to inter-observer disagreement, image re-processing and effect of image filters. Results: Significant intra-class correlations were observed between related sets of features. Histogram-based features were least sensitive to variations with respect to resampling and quantization (78% and 89% of features, respectively, exceeding the highest reproducibility criterion). Global robustness was adversely affected by image filtering and differences in manual delineation. Shape metrics were highly sensitive to inter-observer variability, but were insensitive to differences in image pre-processing. Textural features exhibited the highest sensitivity to processing steps, for all of the examined options. A number of radiomic features were found to be correlated with tumour volume, which should be a noted as a potential confounder for future radiomics studies. Conclusions: This study showed that radiomic features were differentially affected by specific choices made along the processing pathway. In general, histogram features were least sensitive to details of processing, while shape metrics and textural features are more susceptible to heterogeneity in delineation, slice thickness and filtration. The latter was most likely due to its reliance on subtly nuanced properties within the image. Identification and selection of robust radiomic metrics are essential for development of clinically actionable prognostic models for rectal cancer incorporating image-based biomarkers.