Purpose: To improve the reliability, accuracy, and computational efficiency of tissue classification with multispectral sequences [T1, T2, and proton density (PD)], we developed an automated method for identifying training classes to be used in a discriminant function analysis. We compared it with a supervised operator-dependent method, evaluating its reliability and validity. We also developed a fuzzy (continuous) classification to correct for partial voluming.
Method: Images were obtained on a 1.5 T GE Signa MR scanner using three pulse sequences that were co-registered. Training classes for the discriminant analysis were obtained in two ways. The operator-dependent method involved defining circular ROIs containing 5-15 voxels that represented "pure" samples of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), using a total of 150-300 voxels for each tissue type. The automated method involved selecting a large number of samples of brain tissue with sufficiently low variance and randomly placed throughout the brain ("plugs"), partitioning these samples into GM, WM, and CSF, and minimizing the amount of variance within each partition of samples to optimize its "purity." The purity of the plug was estimated by calculating the variance of 8 voxels in all modalities (T1, T2, and PD). We also compared "sharp" (discrete) measurements (which classified tissue only as GM, WM, or CSF) and "fuzzy" (continuous) measurements (which corrected for partial voluming by weighting the classification based on the mixture of tissue types in each voxel).
Results: Reliability was compared for the operator-dependent and automated methods as well as for the fuzzy versus sharp classification. The automated sharp classifications consistently had the highest interrater and intrarater reliability. Validity was assessed in three ways: reproducibility of measurements when the same individuals were scanned on multiple occasions, sensitivity of the method to detecting changes associated with aging, and agreement between the automated segmentation values and those produced through expert manual segmentation. The sharp automated classification emerged as slightly superior to the other three methods according to each of these validators. Its reproducibility index (intraclass r) was 0.97, 0.98, and 0.98 for total CSF, total GM, and total WM, respectively. Its correlations with age were 0.54, -0.61, and -0.53, respectively. Its percent agreement with the expert manually segmented tissue for the three tissue types was 93, 90, and 94%, respectively.
Conclusion: Automated identification of training classes for discriminant analysis was clearly superior to a method that required operator intervention. A sharp (discrete) classification into three tissue types was also slightly superior to one that used "fuzzy" classification to produce continuous measurements to correct for partial voluming. This multispectral automated discriminant analysis method produces a computationally efficient, reliable, and valid method for classifying brain tissue into GM, WM, and CSF. It corrects some of the problems with reliability and computational inefficiency previously observed for operator-dependent approaches to segmentation.