Rupsa Bhattacharjee1, Mamta Gupta2, Tanu Singh2, Shalini Sharma2, Gaurav Khanna3, Suhail P Parvaze4, Rana Patir5, Sandeep Vaishya5, Sunita Ahlawat3, Anup Singh1,6, Rakesh Kumar Gupta7. 1. Center for Biomedical Engineering, Indian Institute of Technology Delhi, Delhi, India. 2. Department of Radiology and Imaging, Fortis Memorial Research Institute, Sector 44, Opposite Huda City Center, Gurugram, Haryana, 122002, India. 3. SRL Diagnostics, Fortis Memorial Research Institute, Gurugram, India. 4. Philips Health System, Philips Innovation Campus, Bangalore, India. 5. Department of Neurosurgery, Fortis Memorial Research Institute, Gurugram, India. 6. Department of Biomedical Engineering, All India Institute of Medical Sciences, New Delhi, India. 7. Department of Radiology and Imaging, Fortis Memorial Research Institute, Sector 44, Opposite Huda City Center, Gurugram, Haryana, 122002, India. rakeshree1@gmail.com.
Abstract
PURPOSE: Primary objective of this study was to retrospectively evaluate the potential of a range of qualitative and quantitative multiparametric features assessed on T2, post-contrast T1, DWI, DCE-MRI, and susceptibility-weighted-imaging (SWI) in differentiating evenly sampled cohort of primary-central-nervous-system-lymphoma (PCNSL) vs glioblastoma (GB) with pathological validation. METHODS: The study included MRI-data of histopathologically confirmed ninety-five GB and PCNSL patients scanned at 3.0 T MRI. A total of six qualitative features (three from T2 and post-contrast T1, three from SWI: thin-linear-uninterrupted-intra-tumoral-vasculature, broken-intra-tumoral-microvasculature, hemorrhage) were analyzed by three independent radiologists. Ten quantitative features from DWI and DCE-MRI were computed using in-house-developed algorithms. For qualitative features, Cohen's Kappa-interrater-variability-analysis was performed. Z-test and independent t-tests were performed to find significant qualitative and quantitative features respectively. Logistic-regression (LR) classifiers were implemented for evaluating performance of individual and various combinations of features in differentiating PCNSL vs GB. Performance evaluation was done via ROC-analysis. Pathological validation was performed to verify disintegration of vessel walls in GB and rim of viable neoplastic lymphoid cells with angiocentric-pattern in PCNSL. RESULTS: Three qualitative SWI features and four quantitative DCE-MRI features (rCBVcorr, Kep, Ve, and necrosis-volume-percentage) were significantly different (p < 0.05) between PCNSL and GB. Best diagnostic performance was observed with LR classifier using SWI features (AUC-0.99). The inclusion of quantitative features with SWI feature did not improve the differentiation accuracy. CONCLUSIONS: The combination of three qualitative SWI features using LR provided the highest accuracy in differentiating PCNSL and GB. Thin-linear-uninterrupted-intra-tumoral-vasculature in PCNSL and broken-intra-tumoral-microvasculature with hemorrhage in GB are the major contributors to the differentiation.
PURPOSE: Primary objective of this study was to retrospectively evaluate the potential of a range of qualitative and quantitative multiparametric features assessed on T2, post-contrast T1, DWI, DCE-MRI, and susceptibility-weighted-imaging (SWI) in differentiating evenly sampled cohort of primary-central-nervous-system-lymphoma (PCNSL) vs glioblastoma (GB) with pathological validation. METHODS: The study included MRI-data of histopathologically confirmed ninety-five GB and PCNSL patients scanned at 3.0 T MRI. A total of six qualitative features (three from T2 and post-contrast T1, three from SWI: thin-linear-uninterrupted-intra-tumoral-vasculature, broken-intra-tumoral-microvasculature, hemorrhage) were analyzed by three independent radiologists. Ten quantitative features from DWI and DCE-MRI were computed using in-house-developed algorithms. For qualitative features, Cohen's Kappa-interrater-variability-analysis was performed. Z-test and independent t-tests were performed to find significant qualitative and quantitative features respectively. Logistic-regression (LR) classifiers were implemented for evaluating performance of individual and various combinations of features in differentiating PCNSL vs GB. Performance evaluation was done via ROC-analysis. Pathological validation was performed to verify disintegration of vessel walls in GB and rim of viable neoplastic lymphoid cells with angiocentric-pattern in PCNSL. RESULTS: Three qualitative SWI features and four quantitative DCE-MRI features (rCBVcorr, Kep, Ve, and necrosis-volume-percentage) were significantly different (p < 0.05) between PCNSL and GB. Best diagnostic performance was observed with LR classifier using SWI features (AUC-0.99). The inclusion of quantitative features with SWI feature did not improve the differentiation accuracy. CONCLUSIONS: The combination of three qualitative SWI features using LR provided the highest accuracy in differentiating PCNSL and GB. Thin-linear-uninterrupted-intra-tumoral-vasculature in PCNSL and broken-intra-tumoral-microvasculature with hemorrhage in GB are the major contributors to the differentiation.
Authors: Bhaswati Roy; Rakesh K Gupta; Andrew A Maudsley; Rishi Awasthi; Sulaiman Sheriff; Meng Gu; Nuzhat Husain; Sudipta Mohakud; Sanjay Behari; Chandra M Pandey; Ram K S Rathore; Daniel M Spielman; Jeffry R Alger Journal: Neuroradiology Date: 2013-02-02 Impact factor: 2.804