Eur. Radiol. (2001) 11: 137±141 Ó Springer-Verlag 2001
E. A. Moore J. P. Grieve H. R. Jäger
Received: 3 February 2000 Revised: 16 May 2000 Accepted: 16 May 2000
)
E. A. Moore ( ) ´ H. R. Jäger Lysholm Radiological Department, National Hospital for Neurology and Neurosurgery, Queen Square, London WC1N 3BG, UK J. P. Grieve ´ H. R. Jäger University Department of Neurosurgery, Institute of Neurology, London WC1N 3BG, UK
NE UR OR A DI O LOG Y
Robust processing of intracranial CT angiograms for 3D volume rendering
Abstract The goal of this study was to develop a robust and simple technique for processing of cranial CT angiograms (CTA) in the clinical setting. The method described in this paper involves segmentation of the bone, then dilation of the skull by adding three or four layers of voxels. This dilated skull is subtracted from the vessels object on a voxel-by-voxel basis, allowing segmentation and subsequent display of the vessels only. For evaluation of the technique, three groups of operators processed one CTA, and the quality of the 3D views obtained and the times taken were compared. One group was given training by an expert and a ªrecipeº for guidance, the second was given only the ªrecipe,º and the third group consisted of expert operators. All operators were
Introduction Since the introduction of spiral CT in the early 1990 s, its use for CT angiography (CTA) has developed rapidly, with applications for examination of the aorta, renal arteries, hepatic circulation and cranial vessels [1, 2, 3, 4, 5, 6, 7, 8]. Despite poorer spatial resolution and a lack of temporal resolution compared with digital subtraction angiography (DSA), CTA has the major advantage of producing a three-dimensional (3D) image instead of projections. In addition, it is a non-invasive technique and can be completed as part of an outpatient examination, with important socio-economic implications. A key difficulty with cranial CTA is in the processing of the 3D volume for visualisation of the vessels [9]. A
able to produce good or acceptable shaded-surface displays when compared with digital subtraction angiography, within 10 min for experienced users, an average of 17 min for trained operators and 26 min for those using only the recipe sheet. Using a simple scoring system for the appearance of feeding vessels and draining veins, no significant differences were found between the three levels of training and experience. This technique simplifies the processing of CTAs and is quick enough to make such examinations part of a routine clinical service. Key words CT angiography ´ Intracranial arteriovenous malformations ´ Computer-assisted image processing
simple maximum intensity projection (MIP) is unsuitable as the vessels are enclosed by the high-intensity skull. Shaded-surface display (SSD) is a popular method, where the vessels appear as solid reflective objects illuminated by a light source, but simple thresholds are usually inadequate to produce a clear angiogram. Complete separation of the vessels from the skull (segmentation) is necessary which requires skilled interaction by the user and long periods of time at the workstation. While this is acceptable in a research setting, it tends to prevent the use of cranial CTA within a routine clinical service which may include 24-h emergency scanning. We believe that in this situation processing should take no longer than 15 min and as far as possible should be independent of the operator's skill. In this technical note
138
we present a processing technique which meets these criteria and the results of its evaluation with six operators of varying expertise.
be separated from other unwanted tissues by selecting with the cursor over the AVM. At this stage final adjustments to the thresholds could be made to leave a clear 3D render of the vessels. The vessels object could also be merged with the undilated skull object, which sometimes improved the visualisation and allowed appreciation of AVM anatomy in relation to bony structures.
Methods The data described in this paper were chosen at random from a large study of CTAs in 19 patients with intracranial arteriovenous malformations (AVMs). All patients gave informed consent for the CTA study which received approval from the local ethics committee. One examination was used in the evaluation of the processing technique and two additional examinations were used as training data. Randomisation was achieved by assigning a number to each examination, then using a random-number generator to produce numbers between 1 and 19 (Microsoft Excel, Microsoft, Redmond, Wash.). The CTAs were acquired on a Somatom Plus-4 scanner (Siemens Medical, Erlangen, Germany) in spiral mode with the following parameters: beam width 1 mm; table feed 2 mm per tube rotation; rotation time 0.75 s; 140 kV; 111 mA; spiral length 68 or 112 mm (depending on the size of the AVM); and reconstruction increment 0.5 mm. A reference slice was used to monitor a region of interest in the cavernous sinus to detect the arrival of contrast agent (1-mm beam width; 1.5-s tube rotation; 111 mA, 140 kV; sample period 3.4 s) and to automatically trigger the CTA acquisition when a pre-determined threshold was reached. A fixed volume of either 100 or 150 ml (depending on the spiral length) of iodinated contrast agent (Xenetix 350, Guerbet Laboratories, Aulnay-sur-Bois, France), regardless of patient weight, was injected at a rate of 3 ml/s using a power injector (AS 200-CT, Medex, France). The total effective dose equivalent to the patient was estimated to be 0.72 mSv. Processing All processing was carried out on an Advantage Windows workstation (GE Medical Systems, Milwaukee, Wis.) running version 1.2 software. With this software, an ªobjectº is a 3D-connected set of voxels whose Hounsfield unit (HU) values are within given thresholds, and objects can be selected for deletion or retention using a cursor. Up to four ªviewportsº can be open, each working independently on the loaded volume. An initial 3D ªmodelº was built using thresholds of 80±1500 HU: this model therefore included all the contrast-enhanced vessels and bony tissues, as well as the head rest and immobilisation pads. The latter objects were deleted immediately because they were not connected to the patient object. Segmentation of the bone from the model was achieved by applying a threshold of 500±1500 HU and then selecting only the skull as the object of interest. This object was then dilated with two to four iterations using the supplied dilation tool. Each iteration dilated the object using a 3 3 3 kernel, adding only voxels which were part of the original 3D model. In a third viewport, the original model was re-thresholded to leave the vessels and to remove as much bone as possible, with typical thresholds of 80±350 HU. This step reduced the memory requirement for processing and allowed work to continue at a faster pace. The dilated skull object was then subtracted from the vessels object, thus creating a clear separation between the intracranial contents and the soft tissues of the face and scalp. Small groups of voxels (ªfloatersº) of less than 25 mm3 were removed to provide a clearer view of the AVM. Finally, the main vessels could
Evaluation of technique Four radiographers were selected and assigned at random to two groups. All four had previously received training by the manufacturer on the basic use of the workstation and were familiar with the standard processing tools. The first group (group A) were individually trained in the technique described previously by one of the authors (E. A. M.), using two CTA examinations (training data 1 and 2) and a ªrecipeº sheet. They were then asked to process a third examination (the test data) without help. The test data were from a patient with a medium-sized AVM (approximately 5 cm) with multiple feeding vessels from the middle and anterior cerebral arteries and partially overlapping draining veins. The second group (group B) received no extra training and were asked to process the test data with only the recipe sheet for guidance. The test data were also processed by two experienced operators (E. A. M. and J. P. G., group C). Four views of the resulting 3D displays were saved on film; lateral and inferior views of the vessels only, inferior view of the vessels and skull, and right-inferior-oblique view of the vessels and skull. The time taken for processing and filming was recorded for each operator. The film outputs were compared with each other and with digital subtraction angiography (DSA) for the test subject. Feeding arteries and draining veins were identified on the DSA and each CTA was scored for visualisation of each vessel (3 = well seen, 2 = partially seen, 1 = not seen) and for definition of the nidus (3 = good, 2 = moderate, 1 = poor). In each case the reviewer (H. R. J.) had the DSA for direct comparison but was blinded to the identity of the operator.
Results Figure 1 shows a view from the right of the processed test data produced by one of the ªexpertº operators, and for comparison Fig. 2 shows lateral frames from the DSA during both the arterial and venous phases. The average training time and processing times for each group of operators are shown in Table 1, and the results of comparing each processed CTA with the DSA are presented in Table 2. The total score for each operator was tested by multiple linear regression against the variables ªtrainingº (groups A and C) and ªexperienceº (group C only). The regression equation was score = 25.5 + 1.5 training + 1.5 experience with low significance on both the coefficients for ¹trainingª and ¹experienceª (p = 0.667). The mean and standard deviation of the scores for arteries and veins were also calculated and found to be 2.5 0.4 and 2.2 0.7, respectively. These were not significantly different using Student's t-test (p = 0.166).
139
Fig. 1 View from the right of a surface-shaded display produced by an ªexpertº operator from the test data. The more conventional lateral view (with anterior to the left of the image) is less useful in this patient whose arterial venous malformation lies in the left hemisphere
a
Discussion The CTA examinations of the thoracic and abdominal aorta [2, 3], the renal arteries [4, 5] and the hepatic circulation [6] are well established as diagnostic methods. In these applications, either maximum intensity projections (MIPs) or shaded-surface displays (SSDs) are fairly straightforward, as there is little or no surrounding bone and the vessels can easily be separated from the surrounding tissues simply using thresholds. The CTA of the cranial vasculature has been used to study intracranial aneurysms [7, 8] and AVMs [9]. In the case of aneurysms, especially those in the circle of Willis, it is relatively easy to define a threshold and a region of interest within the skull, allowing display by either MIP or SSD. However this approach is much less successful when vessels run close to bony structures [7, 8]. In general, MIPs are difficult to use for cranial CTAs [10], since bone always has the highest intensity of all the tissues, and thus the vessels may be obscured by the skull unless there is editing of the volume. Simple SSDs are more useful, with appropriate thresholds and selected volumes of interest to allow visualisation of the vasculature [8]. The choice of thresholds is critical to the final display and may be a cause of artefacts [11]. We believe that to appreciate fully the structure of an AVM, a complete separation of the vessels from the bone is required.
b
c Fig. 2 a±c Lateral digital subtraction angiography of the test case, viewed from the right. a Anterior cerebral artery (arterial phase), b middle cerebral artery (arterial phase), and c venous phase. The dotted line shows the lower extent of the CTA data. Feeding arteries are indicated by small arrows and draining veins by large arrows
Table 1 Training and processing time for three groups of operators Group
Average training time
Average processing time (min)
Training and worksheet Worksheet only Expert operators
1 h 13 min
17 26 7
Segmentation is a task that traditionally requires considerable skill and patience, as well as time [1]. The obvious approach to separating cranial vessels from the skull would involve a combination of thresholds, region growing and manual editing at boundaries. The latter
140
step is necessary because the range of Hounsfield units for contrast-enhanced blood overlaps with that of bone [5], and requires the user to make many decisions about where the boundary should be. On a data set of 150±200 slices, such a task could take a matter of hours and the result will depend on the user having both manual skills at the workstation and good anatomical knowledge. With more sophisticated tools available on modern workstations, it is possible to avoid much of the decision-making involved in manual segmentation, and to reduce the processing time to approximately 15 min. Our technique uses thresholds to produce a rough separation of the vessels, followed by subtraction of a dilated ªskullº object. This approach has numerous advantages over the slice-by-slice editing described previously. Segmentation of the skull is achieved easily with thresholds and requires no specialist anatomical knowledge to recognise a good separation. Segmentation of the vessels requires some care with the lower threshold, to ensure that small vessels are not lost. The upper threshold now does not need to separate vessels from bone, as this is achieved by subtracting the dilated skull object. This allows the user to set a threshold high en-
ough to include all the contrast-enhanced blood, and the processed results are relatively insensitive to the choice of upper threshold. (In our case, we do use an upper threshold in order to reduce memory requirements and to speed up the remaining processing.) Manual skill with the mouse-driven tools is not necessary, because none of the steps involves drawing boundaries. The use of only one data set to assess the results of the processing technique may appear to be a limitation of this paper, and it could be suggested that all six operators should have processed six or more data sets for the analysis. However, they would all have been ªexperiencedº operators by the time they processed the last data set. During preparation of this paper, all radiographers at our institution were trained in this technique, and anecdotal evidence suggests that most become ªexpertº after working on only four or five data sets. Regression analysis of the operators' scores shows that there was a small improvement in the quality of the SSD with specialist training or experience, with a large reduction in the time taken to complete the processing. Thus, we would have expected the time taken for processing to have progressively reduced when working on six data sets, and the quality of the SSDs to have improved slightly; however, the statistics would no longer reflect three groups of operators with a range of training and experience. Another drawback of our evaluation of the processing technique was the use of film output only, since the standard views were not always optimal for all the vessels identified on the DSA. Interactive rotation of the 3D model would have improved visualisation of several vessels in this case, by showing their relationship to overlying structures. The best combination for accurate reporting of CTAs as part of a clinical service is likely to be a set of standard views on film, with the 3D model also viewed interactively on the workstation. With minor modifications to the processing protocol, it is also possible to display the vessels as an MIP
Table 2 Comparison of CTA results with digital subtraction angiography (DSA). A total of seven feeding arteries and three draining veins were identified on DSA. The table shows the number of
arteries and veins shown by each CTA, divided into categories of well seen, partially or poorly seen, or not seen (see text for description of scoring system)
Fig. 3 Maximum intensity projection of the test data, lateral view from the right
Group
Training and worksheet
Worksheet only
Expert operators
Operator
1
2
3
4
5
6
Feeding arteries (maximum seven) Well seen Partially or poorly seen Not seen Draining veins (maximum three) Well seen Partially or poorly seen Not seen
5 2 ±
4 2 1
4 3 ±
2 4 1
5 2 ±
4 3 ±
2 ±
1 1 1
2 1 ±
1 1 1
2 1 ±
2 1 ±
Nidus definition
2
3
3
1
3
1
141
(Fig. 3). Most groups use SSD for CTAs because it offers clearer depiction of pathological vessels relative to overlying structures. However, it has been reported that MIPs are more sensitive and specific than SSDs for renal artery stenosis [5] and for cerebral aneurysms [12], and MIPs have the advantage over SSDs of retaining information about the attenuation within the vessels [13]. We have also undertaken a systematic comparison of the relative merits of MIPs and SSDs for neurological CTA using the data from all 19 patients [14], similar to the study performed by Tanaka et al. [15], and found that MIPs have particular advantages when patients have had previous glue or coil embolisation. In conclusion, we believe that CTA has enormous potential value as an outpatient procedure for monitor-
ing an AVM's response to partial treatment. It is noninvasive and quick to acquire, and involves a small radiation dose (< 1 mSv) compared with DSA (3±4 mSv for a typical four-vessel cerebral angiogram [16, 17]). In order to realise this potential, reliable and swift processing of the 3D images is essential. The technique described in this paper has been demonstrated to produce reasonable quality views even with untrained users, and with experience, operators can produce high-quality SSDs in under 10 min. Acknowledgements The authors acknowledge the financial support of Guerbet Laboratories for this study. We also thank the radiographers who participated in the processing evaluation, and P. Marsden for calculating the radiation doses.
References 1. Zeman RK, Silverman PM, Vieco PT, Costello P (1995) CT angiography. AJR 165: 1079±1088 2. Rubin GD (1997) Helical CT angiography of the thoracic aorta. J Thorac Imaging 12: 128±149 3. Brink JA (1997) Spiral CT angiography of the abdomen and pelvis: interventional applications. Abdom Imaging 22: 365±372 4. Rubin GD (1996) Spiral (helical) CT of the renal vasculature. Semin Ultrasound CT MR 17: 374±397 5. Rubin GD, Dake MD, Napel S, Jeffrey RB Jr, McDonnell CH, Sommer FG et al. (1994) Spiral CT of renal artery stenosis: comparison of three-dimensional rendering techniques. Radiology 190: 181±189 6. Freeny PC (1997) Helical computed tomography of the liver: techniques, applications and pitfalls. Endoscopy 29: 515±523 7. Strayle-Batra M, Skalej M, Wakhloo AK, Ernemann U, Klier R, Voigt K (1998) Three-dimensional spiral CT angiography in the detection of cerebral aneurysm. Acta Radiol 39: 233±238
8. Liang EY, Chan M, Hsiang JHK, Walkden SB, Poon WS, Lam WWM et al. (1995) Detection and assessment of intracranial aneurysms: value of CT angiography with shaded-surface display. AJR 165: 1497±1502 9. Anzalone N, Scomazzoni F, Strada L, Patay Z, Scotti G (1998) Intracranial vascular malformations. Eur Radiol 8: 685±690 10. Velthuis BK, van Leeuwen MS, Witkamp TD, Boomstra S, Ramos LMP, Rinkel GJE (1997) CT angiography: source images and post-processing techniques in the detection of cerebral aneurysms. AJR 169: 1411±1417 11. Takahashi M, Ashtari M, Popp Z, Patel M, Goldstein J, Maguire M et al. (1997) CT angiography of carotid bifurcation: artefacts and pitfalls in shaded-surface display. AJR 168: 813±817 12. Casey SO, Alberico RA, Ozsvath RR (1997) Operator dependence of cerebral CT angiography in the detection of aneurysms. AJNR 18: 790±792 13. Prokop M, Oh Shin H, Schanz A, Schaefer-Prokop CM (1997) Use of maximum intensity projections in CT angiography: a basic review. Radiographics 17: 433±451
14. Jäger HR, Grieve JP, Moore EA, Kitchen N, Taylor W (1999) Comparisons of CT angiography and intra-arterial digital subtraction angiography in patients with cerebral arteriovenous malformations. Neuroradiology 41: 227±228 15. Tanaka H, Numaguchi Y, Konno S, Shrier DA, Shibata DK, Patel U (1997) Initial experience with helical CT and 3D reconstruction in therapeutic planning of cerebral AVMs: comparison with 3D time-of-flight MRA and digital subtraction angiography. J Comput Assist Tomogr 21: 811±817 16. Marshall NW, Noble J, Faulkner K (1995) Patient and staff dosimetry in neuroradiological procedures. Br J Radiol 68: 495±501 17. Steele HR, Temperton DH (1993) Patient doses received during digital subtraction angiography. Br J Radiol 66: 452±456