AUTOMATED VASARI FEATURE SET REPORTING FOR GLIOMAS IS EffCIENT AND EFFECTIVE

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