Data-driven Grading of Brain Gliomas: A Multiparametric MR Imaging Study

Massimo Caulo, Valentina Panara, Domenico Tortora, Peter A Mattei, Chiara Briganti, Emanuele Pravatà, Simone Salice, Antonio R Cotroneo, Armando Tartaro

Purpose To grade brain gliomas by using a data-driven analysis of multiparametric magnetic resonance (MR) imaging, taking into account the heterogeneity of the lesions at MR imaging, and to compare these results with the most widespread current radiologic reporting methods. Materials and Methods One hundred eighteen patients with histologically confirmed brain gliomas were evaluated retrospectively. Conventional and advanced MR sequences (perfusion-weighted imaging, MR spectroscopy, and diffusion-tensor imaging) were performed. Three evaluations were conducted: semiquantitative (based on conventional and advanced sequences with reported cutoffs), qualitative (exclusively based on conventional MR imaging), and quantitative. For quantitative analysis, four volumes of interest were placed: regions with contrast material enhancement, regions with highest and lowest signal intensity on T2-weighted images, and regions of most restricted diffusivity. Statistical analysis included t test, receiver operating characteristic (ROC) analysis, discriminant function analysis (DFA), leave-one-out cross-validation, and Kendall coefficient of concordance. Results Significant differences were noted in age, relative cerebral blood volume (rCBV) in contrast-enhanced regions (cutoff > 2.59; sensitivity, 80%; specificity, 91%; area under the ROC curve [AUC] = 0.937; P = .0001), areas of lowest signal intensity on T2-weighted images (>2.45, 57%, 97%, 0.852, and P = .0001, respectively), restricted diffusivity regions (>2.61, 54%, 97%, 0.808, and P = .0001, respectively), and choline/creatine ratio in regions with the lowest signal intensity on T2-weighted images (>2.07, 49%, 88%, 0.685, and P = .0007, respectively). DFA that included age; rCBV in contrast-enhanced regions, areas of lowest signal intensity on T2-weighted images, and areas of restricted diffusivity; and choline/creatine ratio in areas with lowest signal intensity on T2-weighted images was used to classify 95% of patients correctly. Quantitative analysis showed a higher concordance with histologic findings than qualitative and semiquantitative methods (P < .0001). Conclusion A quantitative multiparametric MR imaging evaluation that incorporated heterogeneity at MR imaging significantly improved discrimination between low- and high-grade brain gliomas with a very high AUC (ie, 0.95), thus reducing the risk of inappropriate or delayed surgery, respectively. © RSNA, 2014.