Automated Brain Extraction from T2-weighted Magnetic Resonance Images

Abstract

Purpose

To develop and implement an automated and robust technique to extract brain from T2-weighted images.

Materials and Methods

Magnetic resonance imaging (MRI) was performed on 75 adult volunteers to acquire dual fast spin echo (FSE) images with fat-saturation technique on a 3T Philips scanner. Histogram-derived thresholds were derived directly from the original images followed by the application of regional labeling, regional connectivity, and mathematical morphological operations to extract brain from axial late-echo FSE (T2-weighted) images. The proposed technique was evaluated subjectively by an expert and quantitatively using Bland-Altman plot and Jaccard and Dice similarity measures.

Results

Excellent agreement between the extracted brain volumes with the proposed technique and manual stripping by an expert was observed based on Bland-Altman plot and also as assessed by high similarity indices (Jaccard: 0.9825± 0.0045; Dice: 0.9912 ±0.0023).

Conclusion

Brain extraction using proposed automated methodology is robust and the results are reproducible.

Introduction

Extraction of the brain from magnetic resonance images is an important pre-processing step for quantitative image analysis. For instance, brain extraction greatly helps in intensity non-uniformity correction, tissue segmentation, non-linear registration, and atrophy measurements (1–10). Brain extraction is commonly employed in assessing progression of neurological diseases such as Multiple Sclerosis (MS) and Alzheimer’s disease (11–14). Currently, majority of the studies depend on computer-assisted or semi-automated brain extraction techniques and involve varying degrees of operator intervention that is both time consuming and prone to human errors.

The volume of cerebrospinal fluid (CSF) can be used for quantitative evaluation of brain atrophy in various neurological disorders such as traumatic brain injury, and MS (7, 9–10). Normalized CSF (nCSF), a measure of brain atrophy, is found to correlate with expanded disability status scale (EDSS) in MS (9). To obtain accurate estimation of CSF, brain extraction based on dual spin echo, particularly, T2-weighted images (or T2 images, for short) is more appropriate than on T1-weighted images (or T1 images). This is because brain extraction on T1 images excludes extra-cortical CSF (15–19). In addition, dual fast spin echo (FSE) images are more appropriate for the classification of gray matter (GM), white matter (WM), CSF, and lesions in adult diseased brains (5, 7, 20). Some of the available automatic techniques for brain extraction include brain surface extractor (BSE), brain extraction technique (BET), and hybrid watershed algorithm (HWA) (16–18). However, these automated brain extraction techniques are focused on T1 images and may not be suitable for dual FSE images.

Suckling et al. (21) have successfully eliminated extrameningeal tissues in dual FSE by deriving linear scale-space features on the proton-density (PD)-weighted image and applying the Laplacian operator at three different scales to define the convexity/concavity. A first-order Gaussian operator at single scale was used to delineate the cortical areas in the T2 images. This algorithm requires the contrast between GM and WM to be uniform over the whole brain for proper identification of the cortex. The GM-WM contrast need not always be uniform throughout the brain in dual FSE images (4). Rajagopalan et al. (22) have proposed a parameter-based technique to transform T2-images into T1-like images so that any method suitable for T1 images could be employed to extract the brain from the transformed T2 images. The parametric projection identifies intensity groups corresponding to background, extra-cortical CSF, WM, GM, and the tissue class comprising of ventricular CSF/fat. With the parameterization of T2 images, T1-like images were obtained in which both extra-cortical and ventricular CSF appear dark similar to normal T1 images. Finally, the well known techniques, such as BET and BSE techniques were applied to T1-like images. Since, T1-like images resemble T1 images, extra-cortical CSF is also eliminated in the process.

In order to improve the classification of normal tissues, GM and WM along with CSF, Grau et al. (23) proposed an improved watershed transform algorithm which was applied to the multi-channel images consisting of T1 and T2 images to obtain intracranial brain mask prior to tissue classification. Zhuang et al. (19) have proposed a model-based level set algorithm to extract brain from T1 and T2 images. This technique was quantitatively evaluated against BSE and BET techniques on T1 images, but only subjectively evaluated on T2 images in normal elderly subjects. Due to the importance of T2 images for quantifying pathology that includes segmentation and atrophy measurements, extrameningeal tissue removal from these images has gained significant attention (21–27). In general, the amount of CSF present in the brain is underestimated by only considering ventricular CSF and excluding extra-cortical CSF on stripping T1 or T1-like images.

The main problem in stripping T2 images arises because the fatty tissues between brain and skull appear bright on T2 images. This complicates the implementation of automated brain extraction techniques for T2 images. A relatively straightforward way to overcome this problem is to acquire T2 images by saturating the fat (fat-sat)surrounding the brain. This fat-sat technique is available on most commercial scanners. Based on the fat-sat T2 images, we have developed an automated brain extraction (ABE) technique to separate brain from extrameningeal tissues. This fully automated technique employs histogram-derived thresholds, region labeling, region connectivity, mathematical morphological operations, and masking to extract brain from axial late-echo FSE (T2) images. The proposed technique was applied to images acquired on 75 adult volunteers. The performance of ABE was evaluated against the extracted brains obtained by an expert using semi-automated method, considered as the gold standard. We have obtained a close agreement between the brain volumes obtained by ABE and expert as also indicated by Bland-Altman analysis (28). The quantitative assessment was also performed using Jaccard and Dice similarity measures (29, 30). These analyses indicate the excellent performance of the proposed ABE technique.

Materials and Methods

Subjects and MRI-Protocol

Seventy five adult volunteers (41 females and 34 males; age range of 20 to 61 years with a median age of 35 years) were included in the present study. Written informed consent was obtained from all subjects. These studies were approved by our Institutional Committee for the Protection of Human Subjects and are fully HIPAA compliant.

A 3T Philips Intera scanner with a quasar gradient system (Philips Medical Systems, Best, Netherlands) and eight channel head coil was used to obtain MRI of the whole brain (from foramen magnumto vertex). Dual FSE images with fat saturation were acquired with echo times, TE1/TE2 = 8.2 ms /90 ms, and repetition time, TR = 6800 ms. The images were acquired in the axial plane with a slice thickness of 3 mm and are contiguous. The total number of axial slices was 44 with an image matrix of 256 × 256 and a field-of-view of either 240 mm × 240 mm or 256 mm × 256 mm. A SENSE factor of 2 was used for all the scans. The total acquisition time for dual echo FSE was approximately7 minutes.

Automated Brain Extraction

Brain extraction essentially involves the identification and elimination of background and separating brain from non-brain tissues such as nasal structures and orbits. We have automatically obtained histogram-derived thresholds for separating neural from non-neural tissues. These thresholds were then applied to brain images to obtain a rough estimate of the intracranial contents. Following the initial brain estimation, the intracranial contents were obtained using regional labeling, regional connectivity, mathematical morphological operations, and masking.

Automated Identification of histogram-derived thresholds

Figure 1 shows T2 images that were acquired with and without the application of fat-sat technique. The suppression of fatty tissues can be well appreciated on the fat-sat T2 images (Fig. 1, column 1). Additionally, there is a well-defined boundary between the brain and extrameningeal tissues in the posterior regions of the fat-sat T2 images. Note that the extrameningeal tissues are mainly present in the anterior region of the brain without apparent boundary with brain tissues. A threshold applicable to obtain boundaries in the posterior region of the brain may not be applicable to the anterior region and therefore two thresholds were identified and applied separately to the anterior and posterior regions of the brain.

Figure 1.

T2 images with (first column) and without (second column) fat-sat technique acquired on same subject.

The histogram, after the application of boxcar averaging to remove local spikes, is shown in Fig. 2. This histogram is characterized by a deep valley, followed by a peak. These two locations, represented by lower threshold (LT) and higher threshold (HT) respectively, were used to separate the brain from the background and other structures. The deep valley (LT) was obtained by equating the first derivative of the histogram to zero with second derivative as positive. Similarly, peak following deep valley (HT) was obtained by equating the first derivative of the histogram tozero with second derivative as negative.

Figure 2.

Histogram of an original fat saturated T2 image demonstrating the locations of lower threshold, LT and higher threshold, HT.

Identification of Intracranial Contents

The axial images were partitioned into two halves to obtain two compartments, each consisting of anterior and posterior regions of the brains prior to the application of thresholds described above. The higher threshold, HT, was applied to the whole image to exclude the background and to disconnect the non-brain structures such as nasal structure and orbits from the brain on each slice, while keeping the brain region. Since no such non-brain structures are attached to the posterior region in the axial brain images, the lower threshold, LT, was used to exclude only background from this region.

Fig. 3 schematically illustrates the steps for identification of the intracranial contents and is described in detail here. Following the identification of LT and HT in the histogram, the image was initially thresholded at HT to obtain the initial mask image ‘I1’, in which the only voxels that have intensity greater than the threshold, HT are included. The largest component of the image, ‘I1’, was identified as image ‘I2’ using 3-dimensional (3-D) connectivity using six adjacent neighbors. Since the white matter is hypointense on T2 images, thresholding often resulted in blank spaces and/or holes within the brain mask. Morphological closing operations (structuring element of spherical shape and size of 3 voxels) were used to separate the blank spaces and/or holes within the mask from background and were classified as part of brain, resulting in image ‘I3’. The resulting mask, ‘I3’, includes orbits and nasal tissues attached to the frontal region of the brain. Morphological erosion was applied to disconnect these tissues from the intracranial brain region to obtain image ‘I4’. Here, the structuring element used was cuboid with 3 voxels in size.

Figure 3.

Schematic diagram demonstrating various processing steps for automated brain extraction (ABE) technique.

Next, the posterior region on each slice of the original brain image, ‘I’, was thresholded at LT to obtain image ‘I5’, by including the voxels with intensity greater than LT. The largest region was identified on all the slices, represented by ‘I5’, following the application of region labeling algorithm by assigning different labels to each region. Two-dimensional (2-D) connectivity using four neighbors was used to expand the identified section on image ‘I5’ to the anterior region on the image ‘I4’ to obtain image ‘I6’. Here, 2-D connectivity was applied on each slice individually to avoid any connection between brain and non-brain structures from superior and inferior slices. This procedure disconnected nasal and other non-neural tissues from the anteriorsection sof the brain (Fig. 3).

In general, the temporal lobes, and sometimes regions at the level of vertex, are present as disjoint structures. Therefore, the application of 2-D connectivity (image ‘I5’ to image ‘I6’) to the axial images resulted in the loss of disjoint temporal lobe(s) (see Fig. 4a). To retrieve these lost temporal lobe(s), it was assumed that the intracranial brain size increases from foramen magnum to the center of the brain and then decreases towards the vertex. The center of the brain is automatically identified as the slice with the largest number of voxels. Therefore, masking the inferior slice with immediate superior slice from the center of the brain towards foramen magnum retrieves the temporal lobes. This step is demonstrated in Fig. 4a. A similar approach was applied towards the vertex of the brain from the center of the brain as this region sometimes appears as two disconnected islands (Fig. 4b). In this case, the superior slice was masked with immediate inferior slice of the brain towards the vertex. This step does not have any effect if there is a single connected region and/or the temporal lobes remained connected from the previous 2-D connectivity approach.

Figure 4.

(a) An example of retrieval of temporal lobes those were lost during initial 3-D and 2-D connectivity. Step 1: Thresholded at HT, 3-D connectivity and morphological operations; Step 2: Posterior region thresholded at LT, 2-D connectivity and morphological operations; Step 3: Masking of slice ‘m’ from left with slice ‘m+1’ from right; Step 4: Retrieval of temporal lobes; and Step 5: Morphological operations and masking with original image; (b) An example of retrieval of vertex regions that were lost during initial 3-D and 2-D connectivity. Step 1: Thresholded at HT, 3-D connectivity and morphological operations; Step 2: Posterior region thresholded at LT, 2-D connectivity and morphological operations; Step 3: Masking of slice ‘m+1’ from left with slice ‘m’ from right; Step 4: Retrieval of vertex region; and Step 5: Morphological operations and masking with original image. Here, ‘m’ and ‘m+1’ represent consecutive inferior and superior slices of the brain.

Retrieval of the temporal lobes and vertex regions resulted in image ‘I7’ (Fig. 3). The applications of HT and morphological erosion to the anterior region result in brain tissue loss at the boundaries. This lost tissue is retrieved by the application of morphological dilation as demonstrated in image ‘I8’. However, application of LT to the posterior region did not result in any tissue loss at the boundaries. The whole brain mask was obtained with the application of 3-D connectivity to obtain single connected region as brain mask, ‘I9’. Finally, this mask was applied toT2 image to extract brain fromextrameningeal tissues (Fig. 3).

Evaluation

All the images included in this study were extracted by an expert, who has more than eight years experience in neuro MRI, using semi-automated software. This semi-automated technique has user-friendly interface and incorporates various tools, such as thresholds, line draw and region isolation etc., providing an interactive platform for the identification of the intracranial region of the brain. This software also provides an option for the operator to move from one slice to another( 2, 6). In the absence of gold standard, the brains extracted by the expert were considered as gold standard and the corresponding brain volumes were used as reference for qualitative and quantitative evaluations in this study. The volumes of the brains extracted by the expert were quantitatively compared with the volumes of extracted brains by ABE using Bland-Altman analysis (28). In the Bland-Altman analysis, the bias between any two measurements defined as their difference, is plotted against the average of two measurements. In the present study, the bias, defined as the difference between the extracted brain volumes by ABE and extracted brain volumes by the expert, was plotted against the average of these two volumes.

Additionally, the performance of the proposed technique was evaluated against the gold standard using Jaccard and Dice similarity indices (29, 30). Also, the fractions of voxels from brain and non-brain regions were evaluated by measuring sensitivity and specificity. Assuming that “Ref” and “Seg” represent the total intracranial brain volumes obtained by an expert and the automated technique used for the extraction of brains, respectively, Jaccard (JSI) and Dice (DSI) similarity indices, sensitivity and specificity are defined as( 31):

$$\text{JSI}=(\text{Ref}\cap \text{Seg})/(\text{Ref}+\text{Sef}-(\text{Ref}\cap \text{Seg})),$$ [1]

$$\text{DSI}=(2\text{x}(\text{Ref}\cap \text{Seg}))/(\text{Ref}+\text{Seg}),$$ [2]

$$\text{Sensitivity}=(\text{Ref}\cap \text{Seg})/\text{Ref},$$ [3]

$$\text{Specificity}=({(\text{Ref})}^{\prime }\cap {(\text{Seg})}^{\prime })/{(\text{Ref})}^{\prime }.$$ [4]

Here, “(Ref)′” and “(Seg)′” represent the complement of sets “Ref” and “Seg”. Higher sensitivity and specificity, defined as: false positive rate = 1- specificity, and false negative rate = 1- sensitivity, respectively, indicate lower false positive and negative rates.

All the techniques were applied to the images without any pre-processing such as noise reduction and intensity non-uniformity correction. All the software including semi-automated brain extraction interface was written under IDL (Interactive Data Language) and implemented on a Dell PC with a 2.99 GHz processor and 3.25 GB of RAM. Manual stripping of the images took approximately 10 minutes per subject using the software. Twenty images used by the expert for training were excluded from the present study.

Results

Brain Extraction

The proposed ABE technique was applied to 75 fat-saturated T2 images to extract the brains and the results are shown in Fig. 5. The validated extracted brains (gold standard) were also included in this figure for reference. As can be observed qualitatively from this figure, the extracted brain sections with ABE technique resemble closely with those obtained by an expert. This was observed to be true for all the subjects included in this study.

Figure 5.

Various slices of T2 image of MR brain (first column), extracted brain slices obtained with ABE technique (second column). Extracted brain slices identified by an expert are also included in the third column for reference.

Evaluation

The agreement between the proposed ABE and the gold standard on T2 images was assessed by using Bland-Altman method. The Bland-Altman plot is shown in Fig. 6. The performance of ABE technique can be appreciated by its close agreement with the expert extracted brain volumes, as indicated by bias that is contained well within the two standard deviations.

Figure 6.

Bland-Altman plot to compare extracted brain volumes by proposed ABE technique and expert on 75 brains.

Similarity measures were computed using equations (1–4) by considering automated extracted brain as “Seg” and extracted brain obtained by an expert as “Ref”. Table 1 summarizes the average Jaccard and Dice similarity indices, sensitivity, and specificity obtained along with respective range (minimum and maximum values) and standard deviations. Our technique took approximately 10 seconds to complete the brain extraction for one subject. The brain volumes extracted from automated technique is plotted against the volumes obtained by the expert using semi-automated technique indicates a high correlation between two set of brain volumes (Fig. 7; R²= 0.9953).

Table 1.

Average values of Jaccard and Dice similarity indices (JSI and DSI), sensitivity and specificity along with range and standard deviations on the comparison of extracted brains obtained with the application of proposed technique on fat saturated T2 images to extracted brains obtained by an expert (gold standard) on 75 brains

Similarity Measure	Average	Minimum	Maximum	SD
JSI	0.9825	0.9591	0.9871	0.0045
DSI	0.9912	0.9791	0.9935	0.0023
Sensitivity	0.9902	0.9699	0.9951	0.0049
Specificity	0.9983	0.9949	0.9996	0.0007

Figure 7.

Estimated brain volumes from proposed technique is plotted against the volumes obtained by an expert. Pearson’s correlation coefficient is R = 0.9976 (R²= 0.9953) with n = 75.

Discussion

In the present study, we have proposed and implemented a fully automated technique for extracting brain from MRI by exploiting the fat-saturation in the T2 images. The method was implemented on axial brain images acquired on 75 subjects. Since the technique was applied to real MR brain images and in the absence of gold standard, the extracted brains by the expert with the help of user-friendly semi-automated technique are considered as gold standard. In the proposed automated technique, the thresholds required to obtain initial brain mask were derived automatically from the image histograms. The regional characteristics of brain were utilized along with the application of mathematical morphological operations to extract the brains. The complete methodology is fully automated and takes about 10 seconds under the IDL environment to extract the whole brain volume including extra-cortical CSF.

The proposed automated technique was applied to fat-sat T2 images which were qualitatively and quantitatively evaluated against the gold standard obtained by an expert. The results demonstrate excellent agreement between the extracted brain volumes obtained with automated and semi-automated technique attesting to the performance of our technique. Since the technique is applied directly to T2 images, the extracted brains include extra-cortical CSF along with ventricular CSF within intracranial brain unlike T1 or T1-likeimages.

The technique presented here is based on the automated identification of thresholds which makes it robust and eliminates any operator bias. The technique is independent of image resolution and is applicable to images with varying resolution s.

In conclusion, the automated method proposed in this study provides accurate and reproducible brain extraction from T2 images as suggested by our qualitative and quantitative analyses on 75 MR brain images. This greatly helps in automating tissue segmentation in various neurological diseases, such as multiple sclerosis (MS) where T2 images are most commonly used. This technique can play an important role in multi-center clinical trials dealing with large amount of patient image data. To the best of our knowledge, this is first study on brain extraction from fat -sat T2 images.