Introduction

Brain extraction is an important step in brain MR image analysis because it influences the accuracy of structural information. For adult brains, T1-weighted (T1W) images are conventionally used for brain extraction because they have better image contrast than T2-weighted (T2W) images. Consequently, the conventional brain extraction algorithms are mostly optimized for T1W images, making it difficult to obtain accurate infant brains from T2W images. As brain extraction from T2W infant brain images is also required in certain clinical applications, an optimized brain extraction algorithm for T2W infant brain images is presented in this abstract.

Methods

The overall procedure of the proposed method is illustrated in Fig.1. Initially, preprocessing is performed to remove the background noise and to compensate for image inhomogeneity. Then, the center slice located in the middle of the brain volume (Fig.2) is selected to generate an initial mask as follows. First, the image is separated into two regions, i.e., background and brain regions, by using the histogram partitioning¹. By applying the 12-class fuzzy c-means (FCM) thresholding² to the separated brain regions and taking the 3^rd threshold image, the rough brain mask of the center slice is generated. Then, refinement is performed to produce the initial mask. The center slice is selected for the initial mask because it usually contains the largest cross-section of the brain. After generating the initial brain mask, brain extraction is performed for the next slice from center to upward, followed by center to downward directions (Fig.2). For the (i+1)^th slice, a rough brain mask is generated as explained above. As some regions with abnormal intensities can be falsely included or excluded during the FCM thresholding, the previous (i^th) brain mask is used to estimate the shape of the current brain slice and to compensate for the errors. However, the previous brain mask may not be suitable for certain image sets having abrupt shape changes between slices (e.g. due to slice gaps), Thus, an inverted skull mask is additionally utilized as a complementary measures. Finally, to include the brain regions that are not connected to the largest component of the brain, the overlap-test³ between the current ((i+1)^th) and the previous (i^th) brain masks is performed. In addition, the eyeball regions are excluded from the mask by considering the location and size of eyeballs⁴ (Fig.3). After removing the eyeball regions, region filling is conducted to refine the brain mask.

Results

We used 2D and 3D infant T2W images acquired at 3T MRI of different vendors (Philips,Siemens) (Table.1). The images were interpolated and registered to RAS coordinates to have voxel sizes of 1×1×1 mm³ and matrix sizes of 256×256×198. 3D datasets contained images with no structural abnormalities and 2D datasets contained images from normal infants and infants with pathological characteristics, including hypo- and hyper-intensities. To quantitatively compare the proposed method, iBEAT⁵, BET⁶ and manual segmentation, dice ratio was calculated after brain extraction. For demonstration purposes, inter-subject average of each slice was calculated for 2D datasets, where the proposed method showed the highest dice ratio (0.8742±0.23), followed by iBEAT (0.8609±0.23) and BET (0.8352±0.23) (Fig.4a). In 3D datasets, dice coefficients of the whole brain volume was calculated to remove the bias due to registration errors and the proposed method also provided better performance in general (Fig.4b). The average dice ratio was 0.9759±0.01 for the proposed method, 0.9608±0.03 for iBEAT, and 0.9603±0.02 for BET, demonstrating that the proposed method generated more accurate brain masks with less performance deviations across datasets. Although iBEAT generated slightly higher dice coefficients in datasets 9 and 10, it showed a large deviation of performance and significant failures in some dataset (data15). In Fig.4c, error images of three different methods and the manual segmentation were also illustrated with different colors, where the proposed method showed the least amount of errors in all slices.

Discussion & conclusion

The proposed method significantly improved the accuracy of brain extraction in infant T2W images, both in 2D and 3D MR images, by using the FCM thresholding with the spatial information of the brain region in the previous slice and the inverted skull mask. While the characteristics of training and test sets may influence the performance of machine-learning based algorithms (as used in iBEAT), the proposed method can be more generally used because it utilizes the characteristics of the target images themselves. The quantitative analysis showed that the proposed method could be efficiently used to extract brain region from infant brain MR images.

Acknowledgements

No acknowledgement found.