Computer-Assisted Detection in FLAIR and DT neuroimages: automatic segmentation and volume assessment of cerebral gliomas

Computer-Assisted Detection in FLAIR and DT neuroimages: automatic segmentation and volume assessment of cerebral gliomas. Giorgio De Nunzio1,6, Marina Donativi1,6, Antonella Castellano2,6, Gabriella Pastore3,6, Matteo Rucco4,6, Antonella Iadanza2, Marco Riva5, Lorenzo Bello5, Andrea Falini2 (1) Univ. of Salento, Dept. of Mathematics and Physics, and INFN, Lecce (2) U.O. Neuroradiologia, Ospedale San Raffaele e Univ. Vita-Salute, Milano (3) Istituto di Ricerche Cliniche Ecomedica, Centro di Radioterapia e IGRT, Empoli (4) Univ. of Camerino, School of Science and Technology, Computer Science Division, Camerino (5) U.O. Neurochirurgia, Ist. Clinico Humanitas, Univ. di Milano, Milano (6) ADAM srl, Advanced Data Analysis in Medicine, http://adamgroup.it Purpose: tumor cells in cerebral gliomas invade surrounding tissues preferentially along WM tracts, spreading beyond the abnormal area depicted on conventional MR images. Diffusion Tensor Imaging can reveal larger peritumoral abnormalities in gliomas that are not apparent on conventional MRI. We aimed at characterizing pathological vs healthy tissue in FLAIR and DTI datasets by 3D statistical Texture Analysis, developing an automatic segmentation technique for cerebral glioma, hereafter called GlioCAD, especially useful in patient follow-up during chemotherapy, and for preoperative assessment of tumor extension. Methods and materials: thirty-four patients with gliomas were selected. 3T axial 3D-FLAIR, axial 3D-T1w, and DTI (single-shot EPI sequence, b=1000 s/mm2, 32 gradient directions) were acquired. Isotropic and anisotropic maps (FA, MD, p and q) were calculated, and pathological ROIs were manually drawn. 3D texture features were calculated with a sliding window approach in the segmented ROIs and in the contralateral healthy tissue, for CAD-system training. The feature-space dimensionality was reduced by Linear Discriminant Analysis, which allowed tissue classification by simple thresholding. Results: For each map, tumor-classification sensitivity, specificity and ROC curves (0.90≤AUC≤0.97) were calculated, and manual and automatic segmentations were compared by the Jaccard Coefficient, showing good concordance. The CAD system automatically calculated lesion volumes and histograms. With the purpose of allowing remote fruition of GlioCAD, a Graphical User Interface was designed as a plugin for OsiriX, a well-known radiological viewer. Conclusion: GlioCAD is proposed as a new tool, based on statistical textural analysis, for the automatic segmentation and volume assessment of brain gliomas, and for the quantitative analysis of the histograms in the regions of interest.