Tumor cells in cerebral glioma invade surrounding tissues preferentially along white matter tracts, spreading beyond the abnormal area seen on conventional MR images. Diffusion tensor imaging can reveal larger peritumoral abnormalities in gliomas that are not apparent on MRI. Our aim was to characterize pathological vs healthy tissue in DTI datasets by 3D statistical Texture Analysis, developing an automatic segmentation technique (CAD) for cerebral glioma, especially useful in a patient follow-up during chemotherapy, and for preoperative assessment of tumor extension. Fifteen patients with glioma (9 low-grade, 6 high-grade) were selected. 3T MR-DTI consisted of a single-shot EPI sequence (b=1000 s/mm2, 32 gradient directions). Fractional anisotropy (FA), mean diffusivity (MD), p and q maps, were obtained. Manual segmentation of pathological areas was performed on each map. 3D texture analysis was applied with a sliding window approach to the segmented ROIs and to the contralateral healthy tissue, in order to identify discriminating features from the intensity and the gradient histogram, and from the cooccurrence (COM) and the run length matrix (RLM). After determining (according to their Fisher-filter score) the best features for each map, the feature-space dimensionality was reduced by Principal Component Analysis, and a neural-network classifier was trained. Glioma segmentations, performed by tissue classification, were compared with the manual ones. Six patients were employed for training, nine for testing. Classifier sensitivity, specificity and ROC curves were calculated: preliminary results were obtained for the p map (AUC = 0.96, sensitivity and specificity equal to 90%, classification error 10.0%) and FA map (AUC = 0.98, sensitivity and specificity equal to 92.6%, classification error equal to 7.3%). Test images were automatically segmented by tissue classification; manual and automatic segmentations were compared, showing good concordance. Our preliminary results show that this approach could allow objective tumor identification and quantitative measurement, with good accuracy.

Automatic Segmentation of Cerebral Glioma in DT-MR Images by 3D Texture Analysis

DE NUNZIO, Giorgio;DONATIVI, MARINA;
2010-01-01

Abstract

Tumor cells in cerebral glioma invade surrounding tissues preferentially along white matter tracts, spreading beyond the abnormal area seen on conventional MR images. Diffusion tensor imaging can reveal larger peritumoral abnormalities in gliomas that are not apparent on MRI. Our aim was to characterize pathological vs healthy tissue in DTI datasets by 3D statistical Texture Analysis, developing an automatic segmentation technique (CAD) for cerebral glioma, especially useful in a patient follow-up during chemotherapy, and for preoperative assessment of tumor extension. Fifteen patients with glioma (9 low-grade, 6 high-grade) were selected. 3T MR-DTI consisted of a single-shot EPI sequence (b=1000 s/mm2, 32 gradient directions). Fractional anisotropy (FA), mean diffusivity (MD), p and q maps, were obtained. Manual segmentation of pathological areas was performed on each map. 3D texture analysis was applied with a sliding window approach to the segmented ROIs and to the contralateral healthy tissue, in order to identify discriminating features from the intensity and the gradient histogram, and from the cooccurrence (COM) and the run length matrix (RLM). After determining (according to their Fisher-filter score) the best features for each map, the feature-space dimensionality was reduced by Principal Component Analysis, and a neural-network classifier was trained. Glioma segmentations, performed by tissue classification, were compared with the manual ones. Six patients were employed for training, nine for testing. Classifier sensitivity, specificity and ROC curves were calculated: preliminary results were obtained for the p map (AUC = 0.96, sensitivity and specificity equal to 90%, classification error 10.0%) and FA map (AUC = 0.98, sensitivity and specificity equal to 92.6%, classification error equal to 7.3%). Test images were automatically segmented by tissue classification; manual and automatic segmentations were compared, showing good concordance. Our preliminary results show that this approach could allow objective tumor identification and quantitative measurement, with good accuracy.
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11587/339098
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact