《外文翻译-- Delineation of intracerebral hemorrhage from clinical non-enhanced computed tomography images》由会员分享，可在线阅读，更多相关《外文翻译-- Delineation of intracerebral hemorrhage from clinical non-enhanced computed tomography images（4页珍藏版）》请在金锄头文库上搜索。
1、Delineation of intracerebral hemorrhage from clinical non-enhanced computed tomography images Qingmao Hu, Zhijun Chen, Jianhuang Wu, Fucang Jia, Shenzhen Institute of Advanced Integration Technology The Chinese University of Hong Kong Shenzhen, China Leping Chen Radiology Department Shekou Peoples
2、Hospital Shenzhen, China AbstractWe delineated intracerebral hemorrhage (ICH) from clinical non-enhanced computed tomography (NCT) images through large grayscale, large grayscale asymmetry with respect to the midsagittal plane (MSP), and large grayscale local contrast. An adaptive approach is propo
3、sed to determine thresholds of the 3 features and adjust the window width for data conversion. The substantial grayscale variability of ICH for a subject is addressed by finding the bright portion followed by recovering. Those ICH voxels symmetrical to the MSP are recovered, partial volume effects a
4、re compensated and the non-ICH high grayscale regions are excluded. The algorithm is fast, and has been validated (37 cases) to yield a delineation accuracy of 0.915 and qualitatively (altogether 161 cases) with good results. The developed system could be a useful tool to aid quantifying ICH and enh
5、ancing stroke therapy. Keywords-intracerebral hemorrhage; generalized asymmetry; CT; segmentation I. INTRODUCTION ICH is an acute and spontaneous extravasation of blood into the brain parenchyma. It accounts for 10-30% of all stroke admissions to hospital, and leads to catastrophic disability, morbi
6、dity, and a 6 month mortality of 30-50% 1. NCT is the recommended modality for the diagnosis of ICH 2, which can be difficult when the lesion is small or is masked by normal structures or when the reader is inexperienced. Estimation of hematoma volume is very useful in predicting patients clinical c
7、ourse and directing management 3. The technical challenges for automatic delineation of ICH include 1) grayscale overlap between non-ICH and ICH voxels; 2) substantial grayscale variability within an ICH region; 3) models such as brain atlas are difficult to be directly incorporated due to the large
8、 voxel size and pathological nature of the clinical NCT data; 4) the mass effect, deformation of brain tissues and spontaneous movement during scanning. Report on quantifying hemorrhage from NCT is scarce. Loncaric et al 4 employed a semi-automatic method by setting seed points and fixed grayscale t
9、hresholds manually to grow ICH. Chan 5 developed an image analysis system to delineate small acute intracranial hemorrhage, which cannot be extended for non-small ICH due to possible deformation and mass effect. However, it is shown 6 that the system can improve the clinicians performance in detecti
10、ng acute intracranial hemorrhage on NCT. At present, the volume and anatomical localization of ICH are approximated manually from NCT slices. This suffers from being tedious and laborious, requiring readers of NCT to be an expert in neuro-radiology, and inability to control accuracy and to provide a
11、ccurate 3D shape information of the ICH 7. It is our objective to devise a robust, fast and validated image analysis system to delineate ICH from clinical NCT data. II. MATERIALS AND METHODS A. Materials One hundred and sixty-one head NCT studies (119 with and 42 without ICH), were retrospectively r
12、etrieved from 3 hospitals in China. All Studies were acquired with a single detector CT scanner (GE or Siemens Medical Systems). The images were axial, obtained parallel to the orbito-meatal line. The slice thickness is either 5mm or 10mm with the matrix being 512x512. Among the 161 subjects, 107 we
13、re male and 54 were female; the age (5418) ranged from 2 days to 90 years. For the 119 ICH data, 8 were in sub-acute and 111 were in acute phases. B. Methods The coordinate system (XYZ) is: X runs from patients right to left, Y from anterior to posterior, and Z from superior to inferior directions.
14、All the anonymized DICOM (Digital Imaging and Communication in Medicine) NCT images were loaded to a Pentium based PC, and were converted to 8-bit format with default window width (WW). Subsequently, the average grayscale of the detected ICH was automatically checked to judge if the ICH is low in gr
15、ayscale (corresponding to sub-acute ICH). If yes, another round is iterated by decreasing the WW of the data conversion. The algorithm consists of: derivation of the brain; characterization of the ICH in terms of 3 features; initial delineation and refinements. These steps are detailed below. Step 1
16、: data conversion Sponsored by NSFC (60803108, 60703120, 30700165) and One Hundred Talent Program of the Chinese Academy of Sciences. 978-1-4244-4713-8/10/$25.00 2010 IEEEUsually it can be assumed that the CT value of acute hemorrhage is around 80 Hounsfield Unit (HU). From the original DICOM volume
17、 orgDV(x,y,z), the 8-bit data orgVol8(x,y,z) can be calculated through the following formula lcH (3) The threshold for asymmetry map asymTh can be determined in a similar way. For brain voxels with grayscale in the range of iL, iH, the typical asymmetry range asymL, asymH can be found. asymTh should
18、 be larger than asymH, asymTh asymH (4) Step 5: derivation of initial hemorrhage initHem(x, y, z) First a voxel-wise binarization is performed, imposing brightness and asymmetry constraints to get B1(x, y, z), i.e., B1(x,y,z) is 1 when (x, y, z) is a brain voxel, orgVolRaw(x, y, z) is not smaller th
19、an heath, and asym(x, y, z) is not smaller than asymTh. The foreground connected component of B1(x,y,z) is found and its local contrast (defined as the difference between the grayscale average of the region and the grayscale average of those voxels that do not belong to the region but are 8-connecte
20、d neighbors of the region) is calculated. For any foreground connected component of B1(x,y,z), if its local contrast is not less than lcTh, and the volume is not less than a constant Const1, then this component is set as the foreground component of the initial hemorrhage initHem(x,y,z). Other voxels
21、 are all set to 0. Step 6: recovering hemorrhage voxels symmetrical to the MSP There are two possibilities: hemorrhage enters lateral ventricle, or the hemorrhage is around the fourth ventricle. For the second case, the hemorrhage is roughly symmetrical around the MSP and the MSP passes through the
22、hemorrhage. Suppose the axial slice with the maximum brain area is zM, then axial slice zp is found as the smallest z with the brain area being 0.8 times the area at slice zM (zp zM). All the axial slices with zzp are checked to find the bright regions which intersect with the MSP and meet the grays
23、cale and local contrast constraints (haeTh and lcTh). These bright regions are then added as the foreground regions of initHem(x,y,z). Other parts of the hemorrhage symmetrical to the MSP can be found in a different way. As the hemorrhage is mainly on one hemisphere, there must be asymmetrical hemor
24、rhage detected in initHem(x,y,z). The missing regions of hemorrhage can be detected this way: these regions meet the grayscale and local contrast constraints and close to some foreground voxels of initHem(x,y,z) (two regions being close means that at least one voxel of a region could find an 8-neigh
25、bor from the other region). Step 7: refinement This is to refine the initial hemorrhage initHem(x,y,z) to include those hemorrhage voxels on the border of hemorrhage regions and to exclude those bright voxels which are non-hemorrhage voxels such as calcification. Calcification is high in grayscale,
26、small in volume and symmetrical to the MSP. It can be excluded by checking the volume (less than Const2) and the asymmetry in a larger window size (11x11, also including upper and lower axial slices, to be smaller than asymTh). Those less bright hemorrhage voxels around the border of hemorrhage regi
27、ons are recovered through checking the 5x5 neighborhood of initHem(x,y,z)s foreground border voxels by decreasing the grayscale threshold to (haeTh+fcmMean3+fcmSd3)/2. Partial volume effect is compensated by checking 3 voxels along 8 direction of the ICH border voxel. If the middle voxels grayscale
28、is larger than the grayscale average of the first (the ICH border voxel) and third voxel, it is changed to foreground voxel. III. RESULTS The algorithm was implemented in C+ and has been validated against the 161 clinical datasets. Twenty (4 sub-acute and 16 acute phases) of the 119 ICH data and 10
29、of the non-ICH data were used to train the parameters of the algorithm. The parameters were chosen as: lcTh = lcH + 5, asymTh = asymH + 5, Const1=10mm3, Const2 = 150mm3 and were kept unchanged for all the data tested. The dependence of ICH accuracy on these parameters was addressed in Discussion. Fi
30、g. 1 shows one sub-acute ICH axial slice and the corresponding detected ICH. a) b) c) Fig. 1: a sub-acute ICH, with an ICH volume 14.64 ml (WW = 80) (b) being adaptively changed to (c) (WW = 70) to get a more accurate ICH volume of 15.92 ml. In order to quantitatively evaluate the algorithm, a radio
31、logist has randomly selected 37 clinical ICH cases (not including the 30 cases for training parameters) to draw the ground truth ICH (with an ICH volume from 0.30 ml to 160 ml, both acute and sub-acute ICH). The measures are Jaccard similarity index,index, false positive (FPR) or false negative rate
32、s (FNR) 10. For the 161 diversified clinical data, all the 119 ICH cases and 42 normal subjects were successfully identified as with or without ICH. For the 37 data, the proposed algorithm yielded average indexes of 0.9150.032, average FPR of 5.36%, and average FNR of 11.02%. IV. DISCUSSION An adapt
33、ive approach is proposed to determine the three thresholds by combining the three features in a novel way. The grayscale threshold haeTh is calculated from voxels with high grayscale asymmetry. The interesting fact is that here “high grayscale asymmetry” does not need to be accurate and can be deter
34、mined through observation or experiments. Here the high grayscale asymmetry should be larger than the grayscale difference between GM and WM, and could be set as 20. Changing 20 to be within a value in 1825 does not change the ICH delineation! The threshold of asymmetry or local contrast is based on
35、 finding the typical brain grayscales as well as the typical asymmetry or local contrast of these typical brain grayscales. Changing the two thresholds respectively, in asymTh-5, asymTh+5 or lcTh-5, lcTh+5 yields only less than 2% change in ICH. Changing Const1 and Const2 with 10% variation yields l
36、ess than 1% change in ICH. This in turn shows the proposed method is robust. It was found that the main errors are due to 3 factors: ICH voxels entering ventricles which are much less bright tend to be missed by the algorithm to yield large FNR (Fig. 2); partial volume voxles in the next axial slice
37、 may be missed by experts but are detected by the algorithm to yield large FPR; small ICH which yields great error with even several voxels mismatch due to drawing (Fig. 3). Otherwise, the detected ICH and the ground truth agree well in appearance. Fig. 2: ICH enters the ventricle and can become muc
38、h less bright (original on the right) which can be identified by expert (middle, the ground truth) but is hard to be detected by the proposed algorithm (left, the detected ICH). This case yielded an FPR 0.91% and FNR 18.1%, respectively. As can be seen from Fig. 1, ICH in sub-acute stage needs a sma
39、ller WW to maintain a relatively constant local contrast. The parameter 160 is determined from the 20 training data. Even for the sub-acute ICH (2 rounds), the execution time on Pentium 4 is less than 1 minute. There are at least three potential uses for our algorithm. First, it could be used for ac
40、curate quantification of ICH, which is especially important for indication of surgery and comparison among different therapies. Second, it could improve workflow, enabling interpreting physicians to focus their attention on patients with ICH by the computer algorithm, and thus more rapidly evaluate
41、patients in the emergency setting. Third, the ICH, together with the major ventricles in 3D (Fig. 4) can help physicians to understand the 3D information of the ICH for better assessing and therapy, especially for junior physicians with less experience. Fig. 3: very small ICH, the detected (right) l
42、ooked similar to the ground truth (middle) but could have large errors (FPR = 9.09%, FNR=3.03%). Fig. 4: a case with ICH entering ventricle (left) and another case with ICH without entering ventricle (right) which can be seen from the 3D ICH (red) and major ventricles (blue). V. CONCLUSIONS We have
43、developed an image analysis system for delineating ICH from NCT. The novelties include: 1) approaching the ICH by combining high grayscales, high local contrast and high grayscale asymmetry with respect to the MSP, and determining these thresholds as well as window width adaptively; 2) addressing th
44、e substantial variability in grayscales of ICH of a subject by determining the bright portion followed by restoring less bright portion; 3) proposing generalized asymmetry to exclude those bright regions symmetrical to the MSP to enhance robustness of the algorithm. The algorithm has been validated
45、against 161 clinical data from 3 hospitals with varying sizes and positions of ICH. It could be a useful tool to aid classifying stroke types, quantifying ICH and enhancing stroke therapy in real clinical setting. REFERENCES 1 Mayer SA, Rincon F. “Treatment of intracerebral haemorrhage. ” Lancet Neu
46、rology. 4, 662-672 (2005) 2 Pendlebury ST, Rothwell PM. “Stroke: management and prevention.” Medicine. 32(10), 62-68 (2004) 3 Butcher K, Laidlaw J. “Current intracerebral hemorrhage management.” Journal of Clinical Neuroscience. 10(2), 158-167 (2003) 4 Loncaric S, Dhawan AP, Broderick J, Brott T. “3
47、-D image analysis of intra-cerebral brain hemorrhage from digitized CT films.” Computer Methods and Programs in Biomedicine. 46, 207-216 (1995) 5 Chan T. “Computer aided detection of small acute intracranial hemorrhage on computer tomography of brain.” Computerized Medical Imaging and Graphics. 31,
48、285-298 (2007) 6 Chan T, Huang HK. “Effect of a computer-aided diagnosis system on clinicians performance in detection of small acute intracranial hemorrhage on computed tomography.” Academic Radiology. 15(3), 290-299 (2008) 7 Chalela JA, Kidwell CS, Nentwich LM, Luby M, Butman JA, Demchuk AM, Hill
49、MD, Patronas N, Latour L, Warach S. “Magnetic resonance imaging and computed tomography in emergency assessment of patients with suspected acute stroke: a prospective comparison.” Lancet. 369, 293-298 (2007) 8 Hu QM, Qian GY, Aziz A, Nowinski WL. “Segmentation of brain from computed tomography head
50、images.” Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference: 155-1155-4 (2005) 9 Hu QM, Nowinski WL. “A rapid algorithm for robust and automatic extraction of the midsagittal plane of the human cerebrum from neuroimages based on local symmetry and outlier removal.” NeuroImages. 20(4), 2154-2166 (2003) 10 Lee JM, Yoon U, Nam SH, Kim JH, Kim IY, Kim SI. “Evaluation of automated and semi-automated skull-peeling algorithms using similarity index and segmentation error.” Computers in Biology and Medicine. 33(6), 495-507 (2003)
