Automatic Partitioning of Head CTA for enabling Segmentation 1

Srikanth Suryanarayanan, Rakesh Mullick, Yogish Mallya, Vidya Kamath, Nithin Nagaraj Imaging Technologies Lab, GE Global Research, Bangalore, India ABSTRACT

Radiologists perform a CT Angiography procedure to examine vascular structures and associated pathologies such as aneurysms. Volume rendering is used to exploit volumetric capabilities of CT that provides complete interactive 3-D visualization. However, bone forms an occluding structure and must be segmented out. The anatomical complexity of the head creates a major challenge in the segmentation of bone and vessel. An analysis of the head volume reveals varying spatial relationships between vessel and bone that can be separated into three sub-volumes: “proximal”, “middle”, and “distal”. The “proximal” and “distal” sub-volumes contain good spatial separation between bone and vessel (carotid referenced here). Bone and vessel appear contiguous in the “middle” partition that remains the most challenging region for segmentation. The partition algorithm is used to automatically identify these partition locations so that different segmentation methods can be developed for each sub-volume. The partition locations are computed using bone, image entropy, and sinus profiles along with a rule-based method. The algorithm is validated on 21 cases (varying volume sizes, resolution, clinical sites, pathologies) using ground truth identified visually. The algorithm is also computationally efficient, processing a 500+ slice volume in 6 seconds (an impressive 0.01 seconds / slice) that makes it an attractive algorithm for pre-processing large volumes. The partition algorithm is integrated into the segmentation workflow. Fast and simple algorithms are implemented for processing the “proximal” and “distal” partitions. Complex methods are restricted to only the “middle” partition. The partitionenabled segmentation has been successfully tested and results are shown from multiple cases.

1. INTRODUCTION CT Angiography is rapidly becoming the method of choice for the examination of vascular structures and their associated pathologies[2,3]. It’s preferred over conventional X-ray angiography because it avoids the risk of inserting catheters into the blood vessels. Current generation scanners are capable of reconstructing the anatomical volume with sub-millimeter resolution in the matter of a few seconds. CT data reveals excellent contrast between bone and tissue, however blood vessel cannot be separated from the soft tissue. To overcome this problem, a radioopaque contrast dye is injected into the blood stream that elevates intensity of blood vessels and provides sufficient contrast from soft tissue. 3-D Visualization exploits the fact that each CT scan captures the anatomical region as a 3-D volume. Rendering techniques such as volume rendering, surface rendering, and Maximum Intensity Projection (MIP) offer greater interaction to the user compared to the conventional multi-planar reformat (MPR) method of viewing the data (2-D views in axial, sagittal, coronal, or oblique planes) [1,4]. For example, an aneurysm in the head can be detected through either of 2-D or 3-D techniques, however, the 3-D techniques provides additional information such as the orientation of the neck of the aneurysm and its reference to nearby arteries. Such information could assist in the planning of interventional procedures. Separation of bone and vessel is critical for 3-D visualization techniques to make a significant contribution. However, the contrast-enhanced vessels have overlapping intensity distribution with the bone regions. Cortical bone forms an opaque visual barrier and trabecular bone forms a haze. This is further complicated in the head where the anatomy presents a larger challenge. The vessels in the head twist and turn within a small volume and pass through bone. Conventionally, bone region have a specific signature with the cortical bone forming a high intensity outer layer with clear boundaries than enclose a spongy lower intensity trabecular bone region. However, such a distinct bone signature is not observed in the head with the skull resembling a thin hollow structure with fuzzy boundaries.

1

Email: [email protected]

Since separation of bone and vessel is challenging in the Head, several studies have looked at a double CT scan to overcome this problem[11]. A non-contrast scan is performed to extract the bone regions followed by a conventional CTA. The two scans are then registered and the bone region extract from the non-contrast scan is masked out from the CTA. While this technique delivers excellent quality of 3-D visualization it increases both the dose to the patient and the scan time. In the clinical workflow today, segmentation of bone and vessel from CTA remains largely a manual process requiring significant operator time [13]. The present way to segment bone and vascular structures in CTA data is through the use of manual contouring: (a) Technologist marks contours around the Region of Interest (ROI), which is typically large bone regions (to remove) or well demarcated vascular structures (to include) (b) This process is repeated for a sufficient number of slices in the volume until a 3-D bone mask or a vascular structure is generated. A second set of methods involves the deposition of several “seed” points by a point and click process. Typically, one or more seeds are required to be deposited inside every vascular structure and region growing is used to extract the vessel tree. With large complex network such as the Circle of Willis in the head, this process can be time consuming and may not cover all the vessels. Moreover, involvement of human operators creates variations between operators, their style of processing along with the radiologist requirements, and the process itself becomes subjective. To avoid the tedious manual processing, semi-automated and automated segmentation techniques have been experimented. Majeid and Avinash[13] developed a semi-automated technique to separate cortical bone from vessels for 3-D visualization of aneurysms and showed results for the abdomen and head CTA. Udupa et al.[7] have established the use of fuzzy connectedness for extraction of 3-D objects using a set of initial starting points. However, the biggest challenge remains complete automation, computational efficiency, and optimal performance across the entire anatomical volume. Even the most sophisticated segmentation methods require user input to initialize the starting conditions or intervene to correct parameters during the process[6,8,10,12]. For practical application in current clinical workflow the segmentation algorithm needs to be fast to and capable of processing a slice well under 0.5 seconds. Our approach involves a unique “divide and conquer” strategy to partition the Head volume into several sub-volumes. The sub-volumes are created such that they maintain consistent spatial relationship between bone and vessel that allows customized algorithms to be developed for each sub-volume.

2. METHODOLOGY Spatial separation becomes critical when objects cannot be separated by intensity difference alone. Good spatial separation between 2 similar objects allows the use of standard techniques such as connected components for differentiating the two objects (Figure 1). However, when the two objects appear contiguous or attached more complex model driven techniques maybe needed to differentiate them. In medical and other applications, two 3-D objects can lie far apart in a region with large spatial separation and lie next to each other in another region with no separation. Therefore, it becomes critical to identify regions within a volume where they are well separated and regions where they may be connected so that different approaches maybe implemented for segmentation.

L1

L2

L1

L1

L2

Figure 1. Two structures “bridged” to each other at one location and well separated at another. Connected components identify both objects with the same label (L1) when they are bridged.

2.1 Data Acquisition Twenty-one data sets acquired from various configurations of GE Medical Systems multi-detector CT scanners were used in this study. Six data sets were randomly selected for training the partition line search algorithm and the remaining 16 data sets were used for testing. The data sets had varying coverage from the aortic arch to the top of the skull. The data sets were loaded as a series of axial cross-sections obtained in DICOM format. 2.2 Anatomical Landscape Analysis The anatomical landscape of the head is analyzed to determine the partition locations required to separate the volume such that each sub-volume contains a consistent spatial relationship between bone and vessel. Our implementation involves a search in the axial direction although the analysis could be performed along any reformatted axis since CT data is volumetric. The loaded CT data is rearranged such that the profile is organized from the distal end in the neck or shoulders and progresses proximally towards the top of the skull. In the neck region, the common carotid arteries are well separated from bone regions. Moreover, they follow a path along the axis of the neck that is manifested as circular regions in the axial slices and can easily be identified. However, the vertebral arteries wind through the cervical vertebra passing through the transverse foramen making separation challenging. The neck region is called the distal partition where carotids remain separate (Figure 2) and vertebrals wind through bone (Figure 3).

Carotids

Bone

Carotids

Bone

Figure 2. The picture in the left illustrates the clear spatial separation of carotids from the cervical vertebra in 3-D rendered view of the neck region. The middle and the right side picture show a sample cross-section in the coronal and axial planes respectively that clearly shows carotids lying lateral to the cervical vertebra and well posterior to the teeth.

Vertebrals

Bone

Vertebrals

Bone

Figure 3. Vertebral arteries wind through the cervical vertebral transverse foramen as illustrated in the 3-D volume rendering in the left picture (bone faded for clear visualization). The sagittal and axial planes in the middle and right pictures also illustrate the path of the vertebral artery through bone.

At the skull base, the most prominent anatomical change is the internal carotid artery looping into the Circle of Willis. The thin petrous and cavernous bone complex appears contiguous with the internal carotid arteries and present a big challenge for separation of the two objects. The region containing the petrous and the cavernous regions in the skull base is called the middle partition (Figure 4).

Figure 4. The picture on the left illustrates the carotids entering the skull bone complex from the distal direction and the picture on the right shows the carotids emerging out of the cavernous bone complex.

At the proximal end of the skull base, the internal carotid vessels emerge out of the cavernous sinus merging into the Circle of Willis. The Circle of Willis is a 3-D topological ring of blood vessels partially embedded in bone. Several branches supplying blood to the corners of the brain emerge from this network. Proximal to the skull base, the only bone structure is the cranium that is well separated from the vessels that are confined to the interior of the region. The cranial section is termed as the proximal partition that also demonstrates clear separation between bone and vessels (Figure 5).

Figure 5. The axial cross-section in the right picture shows the clear spatial separation between the cranial vessels (highlighted) and the skull bone, which is illustrated in 3-D on the pictures to the left where the skull is stripped to reveal the underlying vessels.

2.3 Energy Profiles The previous section provides the anatomical analysis required to partition the Head volume that creates subvolumes with consistent spatial relationship between bone and vessel. Two partition lines that separate the head into distal, mid, and proximal sub-volumes need to be computed. This is achieved by the use of energy profiles. By definition, “energy” of an image is a measure of a selected feature and its distribution in the image. This feature can be modality dependent or modality independent. For CT applications it is desirable to choose bone or air gaps (sinus) as modality dependent features due to their consistent appearance across the volume and across all CT data sets. The profile corresponding to bone and sinus is

created by accumulating all the voxels classified in each image. This is accomplished rapidly because bone and sinus voxels can be easily identified through their distinct intensity distribution. Cortical and trabecular bone can be clearly separated from soft-tissue using a hard threshold of 176 HU. Similarly, sinus regions, which are predominantly large pockets of air, are manifested as large negative HU values. A parallel definition of “energy” in an image comes from the classical theories of entropy [5]. The first order distribution of intensities in an image becomes a measure of how “busy” or “quiet” an image is. The entropy profile is then generated by computing the first order distribution of voxels classified as belonging to bone and vascular regions. The entropy based energy profile is a modality independent measure and can be extended to other volumetric medical data such as MR. 2.4 Search for Partition Lines Each energy profile is stored as an array of accumulated voxels with the first element pointing to the distal end of the volume and the last element of the array pointing to the proximal end. Since bone and the entropy based profile for CT creates a similar map, the bone profile is chosen for further analysis to reduce processing time and redundancy. The partition lines are computed by a rule-based method, which is trained using ground truth identified visually. The rule-based method searches the profiles and identifies the partition slices that correlate to the visually identified slice location. Three hierarchical layers are built to independently identify the partition lines. a) At the first layer, the entire image is used to compute the bone profile using a hard threshold of 176HU. b) For the second layer, a bounding box is computed for each image. The bounding box splits each image into two halves by identifying the most anterior and the most posterior bone voxel. Only bone voxels that lie in the anterior half are accumulated for the profile at this layer. c) In the third layer, the bone and vessel pixels lying in the interior of the head are used to generate the profile. A 3-D search for the farthest bone voxel in the anterior-posterior as well as the lateral directions defines this bounding box and is applied to each image as a mask. The three hierarchical bone profiles and the sinus profile are now the inputs to the search algorithm. The hierarchical bone profiles are used to compute the partition line that separates the distal and middle sub-volumes. The following rules are used: a) Compute the index corresponding to the peak value on each profile. b) Starting at the peak index, the profile is searched in either direction c) Search is terminated upon reaching a minimum pre-computed distance from the index corresponding to the maximum energy location. d) From the index identified in step (c), the nearest index point that corresponds to a 10% decrease in the energy profile value from the peak is chosen as the estimated partition line. e) The estimated location is discarded if there is a change in the slope of the profile along the search or if the search reaches the starting index of the array. The rules for searching the sinus profiles are based on anatomical top-down knowledge. The search is organized in the reverse starting from the top of the skull and decrementing the search index distally towards the neck. The search is terminated upon identifying the first positive (and non-zero) energy value. The search is advanced locally with a tight constraint on the increasing energy value. The first index that corresponds to an energy value above a low threshold is identified as the partition line between the middle and proximal sub-volumes. 2.5 Segmentation of Head CTA Sub-Volumes The Head CTA volume is now separated into three sub-volumes along the axial direction using the partition indices computed. Since each sub-volume now contains a consistent spatial relationship between bone and vessel, segmentation algorithms can now be customized for each sub-volume (Figure 6). Figure 8 shows a fast and a simple

filter implemented to process the proximal partition where there is excellent spatial separation between the skull and the blood vessels. A simple connected components based algorithm is developed to identify the largest island (separate region), which corresponds to the skull (only bone region) in the cranial region. This largest region is then masked out to create a bone-free 3-D rendering. The segmentation method used for middle and distal partitions is beyond the scope of this paper. However, the authors would like to comment that the distal partition was also processed easily using a simple “round regions” based filter and the middle partition was processed using a more complex “recursive constrained region growing” method.

3

RESULTS

Table 1 shows the results of the partition locations computed from 6 training data sets and 15 (different) testing sets with comparison to ground truth (maximum error < 5). Ground truth was established through visual inspection of the data sets and identifying the axial locations of the partition lines. Figure 7 shows the profiles for a case and the rule-based mapping of the partition location on these profiles. The partition location that separates the proximal and middle sub-volume is identified using the sinus profile. The partition location that separates middle and distal subvolume is identified using the bone and image entropy profiles. The maximum error (measured in slices along the axial direction) in estimating the lower and top partition line is 4. The algorithm is extremely fast taking only 0.01 seconds per slice (Dual Intel 1.8GHz processor PC, 2GB RAM, Windows, C++/VTK Debug mode) and processing a 500+ slice volume in 6 seconds. The algorithm is implemented as a VTK filter that can be plugged into any image analysis as a pre-processing step. For segmentation of Head CTA data sets, the partition lines are computed and passed on to the next stage in the process where separate segmentation algorithms are employed for each sub-volume. Figure 9 shows rendering of the cerebral vessels in 3-D. The fast algorithm developed for the proximal partition processes at an average of less than 0.1 seconds / slice (figure 8). Figure 10 shows the output from the proximal, middle and distal partitions from 3 different cases after the segmentation of bone. Since the more complicated region growing based approach is confined only to one sub-volume, the overall segmentation is also fast processing a slice under 0.3 seconds and a typical Circle of Willis data set (150-200 slices) in under a minute. The partition algorithm as well as the segmentation have been successfully tested on several cases collected from various clinical sites with varying head anatomical orientations, contrast levels, and, pathologies.

4

DISCUSSION AND CONCLUSION

The study has demonstrated the use of a “divide and conquer” approach for segmentation of Head CTA through the use of energy profile based partitioning. While energy profiles is in concept similar to the topograms that have been in use for sometime, the fundamental definition and the flexibility through modality dependent and modality independent energy measures add novelty. The partition-based segmentation can be easily extended to other regions of the anatomy. For example, in abdominal CTA, bone profile can be used to isolate the pelvic region, which constantly provides a challenge in the separation of the iliac artery from the pelvic bones. A more complex algorithm with dedicated “de-bridging” can be customized for this anatomical region alone. Although the rules described in this paper are specifically designed to compute partition lines in the head, the search logic can be generalized to other anatomical areas as well. The rules are intended to correlate to the visual findings and can be modified following a thorough analysis of the new anatomical landscape. An issue with such a “divide and conquer” approach is the discontinuity at the partition border. These can arise due to small inaccuracies in the algorithm (this study reports an error of up to 5 slices), varying orientation of the patient, and the fundamental problem that the anatomical landscape is inherently different for each person. Our approach to solve this problem involves overlapping the sub-volumes and developing a separate connected components filter that connects the objects segmented in each partition. The study has developed an extremely fast algorithm (0.01 second / slice) to automatically partition the head volume into 3 convenient partitions (proximal, middle, and distal) based on an analysis of the anatomical landscape.

The partition locations that create sub-volumes are computed using a rule-based method that uses bone, entropy, and sinus profiles. The algorithm has been tested on 20 cases and validated against ground truth. The algorithm has been integrated into segmentation of bone and vessels in head CTA. Simple and fast algorithms have been developed for processing the proximal and distal partitions (< 0.25 second / slice). Complex algorithms are now restricted to only the middle partition that is challenging for segmentation. Results are shown from multiple cases for partitionenabled segmentation

REFERENCES 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.

S. Iserhardt-Bauer, P. Hastreiter, B. Tomandl, N. Kostner, M. Schempershofe, U. Nissen, T. Ertl , Standardized Analysis of Intercranial Aneurysms Using Digital Video Sequences, Medical Image Computing and Computer-Assisted Intervention, v p 411-417, 2002. J. P. Villablanca, R. Jahan, P. Hooshi, S. Lim, G. Duckwiler, A. Patel, J Sayre, N Martin, J. Frazee, J. Bentson and F. Viñuela, , Detection and Characterization of Very Small Cerebral Aneurysms by Using 2D and 3D Helical CT Angiography, American Journal of Neuro-Radiology, 23, 1187-1198, 2002. Venema HW, Hulsmans FJ, den Heeten GJ, CT angiography of the circle of Willis and intracranial internal carotid arteries: maximum intensity projection with matched mask bone elimination-feasibility study, Radiology 18(3):893-8, 2001. R.C. Tam, C.G. Healey, B. Flak, P. Cahoon, “Volume Rendering of Abdominal Aortic Aneurysms, IEEE Visualization Conference, 1997. T. M. Cover and J. A. Thomas, Elements of Information Theory, John Wiley Publications, 1991. o. Wink, W.J. Niessen, and M.A. Viergever, Fast Delineation and Visualization of Vessels in 3-D Angiographic Images, IEEE Transactions on Medical Imaging, vol 18(4), p 337-346, 2000. J.K. Udupa and S. Samarasekara , Fuzzy Connectedness and Object Definition: Theory, Algorithms, and Applications in Image Segmentation, IEEE Transactions on Graphical Models and Image Processing, vol 58(3), p 246-261, 1996. P.K., Saha, J.M. Abrahams, J.K., Udupa, Automatic bone-free rendering of cerebral aneurysms via 3-D CTA, Proceedings of SPIE Medical Imaging Conference, 4322(3), p 1264-1272, 2001. R. Malladi, R. Kimmel, D. Adalsteinsson, G. Sapiro, V. Caselles, J.A. Sethian , A geometric approach to segmentation and analysis of 3-D Medical images, Proceedings of IEEE MMBIA, p 244-252, 1996. Changjiang Yan, Shoji Hirano, Yutaka Hata, Extraction of Blood Vessel in CT Angiographic Image aided by Fuzzy Logic, Proceedings of ICSP, p926-929, 2000. A. Bani-Hashemi, A. Krishnan, and S. Samaddar, Warped Matching for Digital Subtraction of CT-Angiography Studies, Proceedings of SPIE Medical Imaging Conference, v 2710, p 428-437, 1996. S. Shiffman, G. Rubin, and S. Napel, Automated editing of computed tomography sections for visualization of vasculature, Proceedings of SPIE Medical Imaging Conference, v 2707, p 140-151, 1996. Alyassin, A.M. and Avinash, G. Semi-automated bone removal technique from CT Angiography data, Proceedings of SPIE Medical Imaging Conference 2000.

Size

Resolution (mm)

Visual LPL1

Auto LPL1

Visual TPL2

Auto TPL2

Time (total)

512 x 512 x 86 512 x 512 x 203 512 x 512 x 115 512 x 512 x 138 512 x 512 x 119 512 x 512 x 133 512 x 512 x 224 512 x 512 x 138 512 x 512 x 135 512 x 512 x 154 512 x 512 x 126 512 x 512 x 86 512 x 512 x 206 512 x 512 x 176 512 x 512 x 133 512 x 512 x 378 512 x 512 x 505 512 x 512 x 138 512 x 512 x 99 512 x 512 x 157 512 x 512 x 158

.35 x .35 x .6 .49 x .49 x .6 .49 x .49 x .8 .49 x .49 x .49 .49 x .49 x .49 .43 x .43 x .43 .35 x .35 x .63 .39 x .39 x 1.25 .49 x .49 x .6 .43 x .43 x .62 .43 x .43 x .6 .49 x .49 x .6 .35 x .35 x .63 .49 x .49 x .6 .49 x .49 x .6 .49 x .49 x .6 .49 x .49 x .6 .49 x .49 x .5 .49 x .49 x .5 .31 x .31 x .62 .49 x .49 x .63

12 25 18 44 19 22 46 85 50 1 36 4 24 20 24 258 409 14 20 54 41

13 22 19 44 16 19 46 82 49 1 33 2 23 24 21 258 408 13 16 53 38

56 72 70 95 70 67 87 116 110 44 90 40 83 53 90 307 464 70 79 102 88

56 72 68 98 70 67 84 114 110 46 94 37 81 52 94 308 468 66 79 102 90

<1 2 1 1 1 1 2 1 1 1 1 <1 2 2 1 4 6 1 1 1 2

Case

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Table 1. The first six rows represent data sets used for training the partition search algorithm. The bottom sixteen rows show results from testing the algorithm.

Figure 6. The concept of energy profile based partition of the Head volume into 3 sub-volumes is illustrated. Based on the spatial separation between bone and vessel simple and fast extraction algorithms are applied to the proximal and distal sub-volumes and a more complex region growing based approach is restricted to the middle sub-volume.

Entropy

Bone

Sinus lower

Neck

upper

Cranium

Figure 7. The overlay of the two partition line locations estimated over the bone, entropy, and sinus profiles

DICOM data

Partition

Top Partition: Select All Voxels > 200HU

Connected Components

Subtract

Bone Mask

Dilate 3x3x1

Select Largest Region (Volume)

Cerebral Vessels

Figure 8. The flow chart shows a simple and fast method to identify and mask bone voxels after the proximal partition has been created

Figure 9. Results are show for a case at a 2-D slice level and for the entire volume where bone is identified and masked to reveal the underlying cerebral vasculature.

Figure 10. Results are shown for three different test cases where different segmentation algorithms were applied to each subvolume.