Novel solutions toward high accuracy automatic brain tissue classification in young children
Accurate automatic classification of major tissue classes and the cerebrospinal fluid in pediatric MR scans of early childhood brains remains a challenge. A poor and highly variable grey matter and white matter contrast on T1-weighted MR scans of developing brains complicates the automatic categorization of voxels into major tissue classes using state-of-the-art classification methods (Partial Volume Estimation). Varying intensities across brain tissues and possible tissue artifacts further contribute to misclassification. In order to improve the accuracy of automatic detection of major tissue types and the cerebrospinal fluid in infant's brains within the age range from 10 days to 4.5 years, we propose a new classification method based on Kernel Fisher Discriminant Analysis (KFDA) for pattern recognition, combined with an objective structural similarity index (SSIM) for perceptual image quality assessment. The proposed method performs an optimal partitioning of the image domain into subdomains having different average intensity values and relatively homogeneous tissue intensity. In the KFDA-based framework, a complex non-linear structure of grey matter, white matter and cerebrospinal fluid intensity clusters in a 3D (T1W, T2W, PDW)-space is exploited to find an accurate classification. Based on Computer Vision hypothesis that the Human Visual System is an optimal structural information extractor, the SSIM finds a new role in the evaluation of the quality of classification. A comparison with the state-of-the-art Partial Volume Estimation method using the SSIM index demonstrates superior performance of the local KFDA-based algorithm in low contrast subdomains and more accurate detection of grey matter, white matter, and cerebrospinal fluid patterns in the brain volume.