Radiomics features to distinguish glioblastoma from primary central nervous system lymphoma on multi-parametric MRI

• Published in: Neuroradiology Vol. 60; no. 12; pp. 1297 - 1305
• Main Authors: Kim, Yikyung, Cho, Hwan-ho, Kim, Sung Tae, Park, Hyunjin, Nam, Dohyun, Kong, Doo-Sik
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.12.2018; Springer Nature B.V
• Subjects: Adult; Aged; Aged, 80 and over; Brain cancer; Brain Neoplasms - diagnostic imaging; Brain Neoplasms - pathology; Central nervous system; Contrast Media; Diagnosis, Differential; Diagnostic Neuroradiology; Diffusion Magnetic Resonance Imaging; Edema; Feasibility Studies; Feature extraction; Female; Glioblastoma; Glioblastoma - diagnostic imaging; Glioblastoma - pathology; Humans; Imaging; Informed consent; Lymphoma; Lymphoma - diagnostic imaging; Lymphoma - pathology; Magnetic resonance imaging; Magnetic Resonance Imaging - methods; Male; Medical diagnosis; Medical imaging; Medicine; Medicine & Public Health; Middle Aged; Model testing; Nervous system; Neurology; Neuroradiology; Neurosciences; Neurosurgery; Radiology; Radiomics; Regression analysis; Retrospective Studies; Tumors; Diagnosis; Magnetic resonance imaging; Lymphoma; Glioblastoma; Machine learning
• ISSN: 0028-3940; 1432-1920; 1432-1920
• DOI: 10.1007/s00234-018-2091-4

Purpose To determine the feasibility of using high dimensional computer-extracted features, known as radiomics features, in differentiating primary central nervous system lymphoma (PCNSL) from glioblastoma on multi-parametric MR imaging including diffusion-weighted imaging. Methods Retrospective evaluation of data was approved by the local ethics committee and informed consent was waived. A total of 143 patients (two independent cohorts for discovery [ n  = 86; glioblastoma = 49, PCNSL = 37] and validation [ n  = 57; glioblastoma = 29, PCNSL = 28]) with newly diagnosed glioblastoma and PCNSL were subjected to radiomics analysis using the multi-parametric MRI (contrast-enhanced T1-weighted imaging, T2-weighted imaging, and diffusion-weighted imaging). Radiomics analyses were performed for two types of regions of interest (ROI) covering contrast-enhancing tumor and whole (enhancing or non-enhancing) tumor plus peritumoral edema. A total of 127 radiomics features were calculated. Feature selection was performed to identify the most discriminating features for every MR image in the discovery cohort. The identified features were used to calculate radiomics scores, which were later used in logistic regression to distinguish between PCNSL and glioblastoma. The classification model was further tested on the independent validation cohort. Results Fifteen features were selected as significant features in the discovery cohort. Using the identified features and calculated radiomics scores, the logistic regression-based classifier yielded an area under the curve (AUC) of 0.979, sensitivity of 0.938, and specificity of 0.944 in the discovery cohort to distinguish between glioblastoma and PCNSL. A similarly high rate of performance was observed in the validation cohort (AUC = 0.956). Conclusions Radiomics features derived from multi-parametric MRI can be used to differentiate PCNSL from glioblastoma effectively.