Abstract
AIMS
The quantitative evaluation of gliomas using the VASARI MRI feature sets aims to facilitate consistent radiological reporting of gliomas but rarely done in clinical practice due to user variability in addition to technical and time constraints. We questioned whether deep-learning driven systems for automated VASARI feature derivation might resolve these issues, with comparison to neuroradiologist ground truths.
METHOD
VASARI imaging features were derived for 125 gliomas, both manually and using the automated model. Tumour segmentation models were constructed using 1251 patients from RSNA-BRATS with nnU-Net, and a processing pipeline developed within Python. The following features were compared: tumour location (F1); side of lesion (F2); proportion enhancing (F5); proportion non-enhancing (nCET) (F6); multifocal (F9); proportion edema (F14); ependymal extension (F19) and deep white matter (WM) invasion (F21).
RESULTS
The out-of-sample Dice performance for the tumour segmentation model was 0.945. Automated VASARI feature sets could be derived at a rate of under 0.5 seconds per patient. There was high concordance between manual and automatic segmentation in determining side of lesion (F2, 91.2%), ependymal extension (F19, 86.4%), anatomical tumour location (F1, 77.6%), deep WM invasion (F21, 65.6%) and proportion of enhancing tumour (F5, 54.4%; r=0.377, p<0.001). Automated segmentation overestimated proportions of edema (F14, 92.8%) and nCET (F6, 86.4%).
CONCLUSIONS
Automated VASARI feature set generation provides quick and accurate predictions however, proportions of oedema and non-contrast tumour were frequently over-estimated. Future empirical work will focus on optimization and further systematic evaluation of this automated pipeline for more accurate quantitative radiology reporting and corresponding clinical decision-making.