Adding automated decision-tree models to multiparametric MRI for parotid tumours improves clinical performance.

PURPOSE: Therapeutic management of parotid gland tumours depends on their histological type. To aid its characterisation, we sought to develop automated decision-tree models based on multiparametric magnetic resonance imaging (MRI) parameters and to evaluate their added diagnostic value compared with morphological sequences. METHODS: 206 MRIs from 206 patients with histologically proven parotid gland tumours were included from January 2009 to January 2018. Multiparametric MRI findings (including parameters derived from diffusion-weighted imaging [DWI] and dynamic contrast-enhanced [DCE]) were used to build predictive classification and regression tree (CART) models for each histological type. All MRIs were read twice: first, based on morphological sequence findings only, and second, with the addition of multiparametric sequences and CART findings. The diagnostic performance between these two readings was compared using ROC curves. RESULTS: Compared to morphological sequences alone, the addition of multiparametric analysis significantly increased the diagnostic performance for all histological types (p < 0.001 to p = 0.011), except for lymphomas, where the increase was not significant (AUC 1.00 vs. 0.99, p = 0.066). ADCmean was the best parameter to identify pleomorphic adenomas, carcinomas and lymphomas with respective cut-offs of 1.292 x 10(-3) mm(2)/s, 1.181 x 10(-3) mm(2)/s and 0.611 x 10(-3) mm(2)/s, respectively. x 10(-3) mm(2)/s. The mean extracellular-extravascular space coefficient was the best parameter to Warthin tumours from the others, with a cut-off of 0.07. CONCLUSIONS: The addition of decision tree prediction models based on multiparametric sequences improves the non-invasive diagnostic performance of parotid gland tumours. ADC and extracellular-extravascular space coefficient are the two best parameters for decision making.