348 Multi-Center MRI Radiomic Texture Analysis Pre-Operatively Discriminates Between Primary Central Nervous System Lymphoma, Glioblastoma, and Brain Metastasis. (1st April 2022)

• Record Type: Journal Article
• Authors: Pease, Matthew; Ak, Murat; Ayoub, Mira; Mamindla, Priyadarshini; Kamal, Abdallah; Peddagangireddy, Vishal; Zinn, Pascal O.; Moron, Fanny; Colen, Rivka R.
• Is Part Of: Neurosurgery. Volume 68(2022)Supplement 1
• Page Start: 81
• Page End: 82
• Publication Date: 2022-04-01
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1227/NEU.0000000000001880_348
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract : INTRODUCTION: Primary Central Nervous System Lymphoma (PCNSL), glioblastoma (GBM), and brain metastasis (BM) collectively account for almost 75% of malignant brain tumors. Due to their similar radiographic appearance, the current standard of care is to perform an invasive biopsy for histological diagnosis. Radiomic approaches may allow for non-invasive identification of tumor histology to guide patient care. METHODS: We retrospectively identified 632 patients from three institutions with untreated solitary, pathology confirmed cases of PCNSL (133), GBM (129), and breast or lung BM (370). For each tumor, we segmented edema, contrast enhancement, necrosis, and contralateral white matter (cWM) regions of interest (ROI) for T2 FLAIR and T1 post-contrast sequences. A total of 205 texture features were extracted from each sequence's ROI including 10 first-order features and 195 second-order features using gray-level co-occurrence matrices. We divided the 195 second level features by the volumes of each ROI to generate another 195 volume-independent features and 2, 400 features total. We removed all highly correlated feautres, performed feature selection with Least Absolute Shrinkage and Selection Operator (LASSO), and built a model using gradient boosting. We split the cohort into 80% training and 20% testing with 5-fold cross-validation. We report area under the curve (AUC) for classifying each tumor type against all others (i.e., BM versus PCNSL and GBM). RESULTS: LASSO identified 173 independent features, of which 35 most relevant were used for model building. Our three-classification model robustly differentiated (p < 0.001) between PCNSL (AUC 0.95), GBM (AUC 0.93), and BM (AUC 0.93). Our accuracy was 84% and we achieved favorable sensitivity (PCNS 0.74; GBM 0.73; BM 0.91) and specificity (PCNSL 0.97; GBM 0.92; BM 0.83). With a specificity of 1 (i.e., never incorrectly classifying a tumor subtype), the sensitivity for PCNSL and BM was over 40% and approached 80% for GBM. CONCLUSION: MRI radiomic analysis may achieve enough classification accuracy to allow for avoidance of intracranial biopsy in select patients and differentiate between PCNSL, GBM, and BM.