Method and device for detecting bright brain regions from computed tomography images
A method and devices are disclosed to detect bright brain regions (BBRs) from clinical non-enhanced computed tomography images through large grayscale, large grayscale asymmetry with respect to the midsagittal plane (MSP), and large grayscale local contrast. An adaptive approach is disclosed to determine thresholds of the 3 features and adjust the window width for data conversion. The substantial grayscale variability of BBRs for a subject is addressed by finding the bright portion followed by recovering. Those BBR voxels symmetrical to the MSP are recovered, partial volume effects are compensated and the high grayscale regions which may not correspond to intracerebral hemorrhage are excluded. The disclosed method and system could be a useful tool to aid classifying stroke types, quantifying intracerebral hemorrhage and enhancing stroke therapy.
1. A method for detecting the bright brain regions from computed tomography images, comprising:
loading computed tomography image data;
deriving images of a skull;
deriving brain images;
converting the brain images to lower bit images data;
calculating a generalized asymmetry map with respect to the midsagittal plane (MSP);
calculating a grayscale threshold, a threshold of a local contrast, a threshold of asymmetry, and a threshold of partial volume voxels for bright brain regions (BBRs);
deriving initial bright brain regions (BBRs);
recovering bright brain voxels symmetrical to the midsagittal plane (MSP);
refining the initial bright brain regions (BBRs);
checking if detected bright brain regions (BBRs) is different from an average grayscale,
increasing or decreasing a window width being increased or decreased when the detected bright brain regions (BBRs) is different from the average grayscale, then and going back to the converting step; and
after checking to ensure that the bright brain regions (BBRs) match the average grayscale, outputting the bright brain regions (BBRs).
2. The method as claimed in claim 1 , the step of deriving the skull images, further comprising:
deriving the skull from the computed tomography images in which a threshold has been calculated, in order to contain each of the voxels with a computed tomography value greater than a constant corresponding to the skull.
3. The method as claimed in claim 1 , wherein said step of deriving the brain images further comprises:
applying Fuzzy c-means clustering to the computed tomography images in which a threshold has been calculated, in order to classify the volumetric data into clusters with ascending grayscales to get low and high thresholds to remove non-brain tissues;
achieving the binarization through removing non-brain voxels to derive B 1 (x, y, z), i.e.,
B
1
(
x
,
y
,
z
)
=
{
1
,
low
threshold
<
orgDV
(
x
,
y
,
z
)
<
high
threshold
0
,
otherwise
;
applying a morphological opening with a rectangular structuring element with a side of 4 mm to B 1 (x, y, z) in order to break the connection between the brain and non-brain tissues to get B 2 (x, y, z);
finding a maximum foreground connected component of B 2 (x, y, z), holes of which are then filled to get B 3 (x, y, z); and
checking neighboring axial slices of B 3 (x, y, z) to exclude isolated foreground regions of B 3 (x, y, z) to get the eventual brain mask (Brain) (x, y, z).
4. The method as claimed in claim 3 , wherein during said the step of deriving the brain images, the low threshold is set as a maximum intensity of a first one of the clusters, and the high threshold is set as a constant corresponding to the skull.
5. The method as claimed in claim 4 , wherein during said step of deriving the brain images, each of the isolated foreground regions at the axial slice is a foreground region having fewer than half of the voxels (x, y, z) than in corresponding f ones of the voxels at its lower and upper axial slices.
6. The method as claimed in claim 1 , wherein during said loading step, the brain images of 16-bit image data are converted to 8-bit data.
7. The method as claimed in claim 6 , wherein during the converting step, the 8-bit brain data is formulated for both processing and monitoring with a fixed window width through the following formula:
orgVolRaw
(
x
,
y
,
z
)
=
{
0
,
Brain
(
x
,
y
,
z
)
<
0
255
×
Brain
(
x
,
y
,
z
)
window
width
,
0
≤
Brain
(
x
,
y
,
z
)
≤
window
width
255
,
Brain
(
x
,
y
,
z
)
>
window
width
wherein the window width is a constant and is changed adaptively according to a brightness of the detected bright brain regions (BBRs).
8. The method as claimed in claim 1 , the calculating step of the generalized asymmetry map further comprising:
deriving a head mask through binarization of the computed tomography images;
extracting the midsagittal plane (MSP) by deriving from the computed tomography images through a formula as follows, with the window width being a constant corresponding to the skull:
orgVolRaw
(
x
,
y
,
z
)
=
{
0
,
Brain
(
x
,
y
,
z
)
<
0
255
×
Brain
(
x
,
y
,
z
)
window
width
,
0
≤
Brain
(
x
,
y
,
z
)
≤
window
width
255
,
Brain
(
x
,
y
,
z
)
>
window
width
;
defining the local symmetry as a sum of a grayscale difference of a plurality of pairs of searching line segments, and
removing in order to handle an abnormal grayscale distribution in axial slices; and
wherein the generalized asymmetry map is denoted as asym (x, y, z) and is calculated as:
asym
(
x
,
y
,
z
)
=
{
orgVolRaw
(
x
,
y
,
z
)
-
max
(
reg
(
x
′
,
y
′
,
z
)
)
,
orgVolRaw
(
x
,
y
,
z
)
>
max
(
reg
(
x
′
,
y
′
,
z
)
)
0
,
otherwise
wherein (x′, y′, z) is the mirrored point of (x, y, z) with respect to the intersection line at axial slice, max (reg (x′, y′, z)) is a maximum grayscale of orgVolRaw(x,y,z) within a neighborhood centered at (x′,y′,z).
9. The method as claimed in claim 8 wherein during the step of deriving the head mask, any of the voxels with a grayscale not less than T low is set to 1 and 0 otherwise; and
a maximum connected component of the binary image is found and holes within a maximum foreground connected component are filled to form the head mask.
10. The method as claimed in claim 8 , wherein between the defining step and the step of calculating the generalized asymmetry map, the calculating step further comprising:
approximating each of the head mask as an ellipse with the long axis being taken as an approximated symmetry line segment to be used for approximating the midsagittal plane (MSP) through outlier removal, if a number of inliers is less than 6; and
denoting an equation of the midsagittal plane (MSP) as a x+by+cz+d=0,
wherein a, b, c, d are calculated parameters with a>0, a 2 +b 2 +c 2 =1, and x, y, z, and d are all in mm.
11. The method as claimed in claim 8 wherein during the calculating step of the generalized asymmetry map, max (reg(x′, y′, z)) can be obtained by calculating the grayscale average and a standard deviation of the axial slice and to obtain the sum as max (reg(x′, y′, z)).
12. The method as claimed in claim 8 , the calculating step of calculating the thresholds, further comprising:
supposing the grayscale mean and a standard deviation of the voxels as avg 1 and sd 1 , and calculating the grayscale threshold for the bright brain regions (BBRs) as a grayscale threshold=avg 1 −sd 1 ;
basing the threshold of the local contrast and the threshold of asymmetry for the bright brain regions (BBRs on an analysis of histograms, to be larger than of typical brain voxels, and
further comprising:
identifying the typical brain voxels by performing steps as follows:
supposing a grayscale histogram of each of the brain voxels is hist brain (i),
supposing a grayscale having the maximum frequency of i most ,
supposing a smallest grayscale having half of the maximum frequency of i L , and
supposing a largest grayscale having half of the maximum frequency of i H ;
obtaining [0, i L ), [i L , i H ] and (i H , 255) as low, typical and high grayscales of the brain, respectively;
calculating the local contrast for any brain voxel (x, y, z) with the grayscale in the range of [i L , i H ], by calculating the local contrast as orgVolRaw (x, y, z) minus a minimum grayscale as compared to eight neighbors;
calculating a histogram of each of the local contrasts and a typical range of the local contrasts [Ic L , Ic H ], the threshold of the local contrast being larger than contrast Ic H ;
determining a threshold for an asymmetry map for brain voxels with the grayscale in the range of [i L , i H ], by finding a typical asymmetry range [asym L , asym H ],
the threshold of the asymmetry range being larger than asym H .
13. The method as claimed in claim 8 , the step of deriving initial bright brain regions (BBRs) further comprising:
performing a voxel-wise binarization,
imposing brightness and asymmetry constraints to get B 4 (x, y, z);
for any brain voxel setting (x, y, z), B 4 (x, y, z) set to 1 only when orgVolRaw (x,y,z) is not less than haeTh, and asym(x,y,z) is not less than asymTh, otherwise setting B 4 (x,y,z) to a setting of 0;
finding the foreground connected component of B 4 (x, y, z);
defining the local contrast as a difference between the grayscale average of the region and the grayscale average of the voxels that do not belong to the region but are 8-connected neighbors of the region; and
for any foreground connected component of B 4 (x, y, z), if the local contrast is not less than threshold of the local contrast, and the volume is not less than the constant,
setting the foreground connected component as the foreground connected component of the initial bright brain regions (BBRs), and
setting other voxels to the setting of 0.
14. The method as claimed in claim 8 , the recovering step further comprising:
checking the bright brain regions (BBRs) against the grayscale threshold (larger than haeTh), the local contrast (larger than lcTh) and volumes;
supposing the axial slice with a maximum brain area is z M , the axial slice z p being found as the smallest z with the brain area being 0.8 times the area at slice z M if z p >z M ;
checking each of the axial slices with z>z p to find the bright brain regions (BBRs) which intersect with the midsagittal plane (MSP) and meet the grayscale and local contrast constraints;
adding the bright brain regions (BBRs) as the foreground regions;
finding other parts of the bright brain regions (BBRs) symmetrical to the midsagittal plane (MSP) in a different way; and
detecting the missing bright brain regions (BBRs) which meet the grayscale and local contrast constraints and close to some voxels in foreground regions.
15. The method as claimed in claim 1 , the refining step further comprising:
including the bright brain voxels on a border of each of the bright brain regions (BBRs);
excluding the bright voxels which may obviously correspond to non-hemorrhage voxels such as calcification;
recovering fewer of the bright brain voxels around the border of bright brain regions (BBRs)
compensating for a partial volume effect; and
excluding the voxels in the foreground regions from the computed tomography images greater than the constant corresponding to the skull.
16. The method as claimed in claim 15 , the including step further comprising:
checking the volume and checking the asymmetry in a larger window size; and
setting the initial bright brain regions (BBRs) to a setting of 0, if the volume of any of the foreground regions in the initial bright brain regions (BBRs) is smaller than the constant and does not have an asymmetry map greater than the threshold of asymmetry.
17. The method as claimed in claim 15 , the recovering step further comprising:
checking a 5×5 neighborhood of foreground border voxels of the initial bright brain regions (BBRs), and
setting the initial bright brain regions (BBRs) to 1, if a background voxel is within a 5×5 neighborhood and the grayscale of the third cluster is larger than (grayscale threshold+grayscale mean of the third cluster+standard deviation of the third cluster)/2, and the voxel contains being at least half of the bright brain voxel, and
obtaining the background voxel as the bright brain voxel.