AUTOMATED METHODS FOR GENERATION OF ARTERIAL AND

VENOUS REFERENCE POINTS FOR CONTRAST-ENHANCED

MAGNETIC RESONANCE ANGIOGRAPHY AND

COMPUTED TOMOGRAPHY

FIELD OF THE INVENTION

The present application relates to medical imaging devices, in general, and to methods, systems, and programs of instructions for the identification of vascular diseases based on image data such as MRI and CT or any other medical imaging procedure and device.

BACKGROUND OF THE INVENTION Vascular occlusive disease is a major public health problem. More than 100,000 patients undergo reconstructive vascular surgery in the United States every year. The diagnosis of this disease requires precise imaging of the vascular pathological areas. It is essential to get an accurate assessment of the vascular system for surgical planning. In some parts of the body, the precise mapping of the arterial vessel patterns, based on MR or CT image data is often obstructed by veins. Artery-vein segmentation (AVS) within 3D or 2D image datasets could significantly assist radiologists and vascular surgeons in accurate diagnosis and appropriate treatment planning.

Current AVS strategies for contrast enhanced angiography

Bolus timing with test bolus. This is the most simple and widely used method for obtaining arterial images without obstruction from venous structures. Since the duration of time for contrast agent (CA) arrival in the organ of interest (vascular circulation time) can vary greatly between patients, a small dose of CA is administered before the actual angiographic scan to measure the patient's specific circulation time [M. R. Prince, T. M. Grist, J. F. Debatin, "3D Contrast MR Angiography", Springer-Verlag (2003)]. The main limitation is that the "arterial phase" (period of time when CA has arrived only in the arteries) for imaging may only be 10-15 seconds prior to the CA arriving in the veins.

This short acquisition window for an arteriogram (arterial map) limits the spatial resolution and the signal-to-noise ratio of the resulting MRA image.

Bolus timing with pre-defined vessel screening. With this method, the scanner itself screens a pre-defined vessel (manually selected by the physician) for the arrival of CA. When the CA arrives at that vessel, the scanner automatically starts the image acquisition for the magnetic resonance angiography (MRA) sequence [M. R. Prince, T. M. Grist, J. F. Debatin, "3D Contrast MR Angiography", Springer-Verlag (2003)]. Similar to the previous method, the short acquisition window for an arteriogram limits the spatial resolution and the signal-to noise ratio.

Bolus chasing. Bolus chasing ("moving table technique") requires the ability to perform multiple 3D acquisitions rapidly, move the patient table quickly and accurately, and administer the CA at constant rate. Higher spatial resolution requires longer acquisition times, which in turn delay time of imaging for all subsequent stations. This method is also very user dependent and consequently the MRA data can be of variable quality. Post-processing tools. The most commonly used post-processing

AVS tool has been a technology called "fuzzy logic" [Dellepiane, S. G. and Novelli, L. and Bruzzo, M. and Antonelli, M., "A Fuzzy Connectivity Tree for Hierarchical Extraction of Venous Structures", Scandinavian Conference on Image Analysis SCIA03, pp. 844-852 (2003)]. This method segments arteries and veins using fuzzy connected object-delineation principles and algorithms. In other words, the method determines that contiguous, similarly enhanced pixels represent an individual vein or artery. When any part of an artery or vein is indicated manually by a physician or MR technologist, the software can reconstruct the whole vessel by following the chain of contiguous pixels. The main disadvantage of this method is that it is not fully automated and requires manual selection of arterial and venous "seeds" from which the algorithm will begin to reconstruct the vessels [Dellepiane, S. G. and Novelli, L. and Bruzzo, M. and Antonelli, M., "A Fuzzy Connectivity Tree for Hierarchical Extraction of Venous Structures", Scandinavian Conference on Image Analysis SCIA03, pp. 844-852 (2003)].

SUMMARY OF THE INVENTION

A method, system, and computer-readable medium for fully automatic (controlled by the software without user interaction) segmentation of arteries and veins based on the time-resolved image data, is provided. The method consists of three major parts, where each one is independently automated.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 shows a splitting of the FOV in the multiple ROIs.

Fig. 2 shows a dataset structure for time-resolved 3D MRA

Fig. 3 shows a dynamic enhancement curve for peak enhancement and arrival time for an individual pixel. Fig. 4 shows a joint time-intensity histogram for automatic generation of reference curves with arterial and venous enhancement patterns.

Fig. 5 shows an arterial (light) and venous (dark) reference curves generated for each of the upper seven ROIs in Fig. 1.

Fig. 6 shows an original high-resolution post-contrast MRA dataset and automatically generated multiple reference pixels of arterial type ("seeds").

Fig. 7 shows a segmented and colored 3D dataset with soft (enhancement dependent) arterial (venous) segments stored in red (blue+green) channels of the RGB dataset (DICOM, etc).

DETAILED DESCRIPTION OF THE INVENTION

Splitting FOV into multiple regions In this method, the whole field of view (FOV, the anatomy/body volume that was imaged) is divided into multiple regions of interest (ROIs) for more accurate processing to account for the delayed arrival of the CA in the more distant parts of the body [M. R. Prince, T. M. Grist, J. F. Debatin, "3D Contrast MR Angiography", Springer -V erlag (2003), Bock M, Schoenberg S. O., Flomer F., Schad L. R., "Separation of arteries and veins in 3D MR angiography using correlation analysis", Magn. Reson. Med. 43, pp. 481-187 (2000), Mazaheri Y,

Carroll T. J., Mistretta C. A., Korosec F. R., Grist T. M., "Vessel segmentation of 3D MR angiography using time resolved acquisition curves", Proc. ISMRM 6th Scientific Meeting, Philadelphia, p. 2181 (1999)]. An example of such a division is shown in Fig. 1. The FOV containing lower extremities is divided into 8 ROIs (vertical division). In other parts of a body horizontal or combined splitting can be used. All calculations are then performed independently in each of the generated ROIs.

Automatic generation of reference curves

For 3D time-resolved acquisition [M. R. Prince, T. M. Grist, J. F. Debatin, "3D Contrast MR Angiography", Springer-Verlag (2003), Bock M, Schoenberg S. O., Flomer F., Schad L. R., "Separation of arteries and veins in 3D MR angiography using correlation analysis", Magn. Reson. Med. 43, pp. 481-187 (2000), Mazaheri Y, Carroll T. J., Mistretta C. A., Korosec F. R., Grist T. M., "Vessel segmentation of 3D MR angiography using time resolved acquisition curves", Proc. ISMRM 6th Scientific Meeting, Philadelphia, p. 2181 (1999)] the image dataset can be presented as a 4D object S(x, y, z, t), where S is the acquired signal intensity for (x, y, z) location at time t. Locations x and y are in- plane coordinates (x = 0, ... ,Nx- 1 ; y = 0, ... ,Ny- 1 ), z is the slice number (z =

0,...,Nz-I), and t - time-point index (t = 0,...,Nt-I), as shown in Fig. 2. Here Nx and Ny represent in-plane matrix (single image dimensions), Nz - number of slices, and Nt - number of dynamic phases [Prince et al. (supra)] including pre- contrast (the minimum number of dynamic phases required is three). For each individual pixel, the dynamic intensity curve is defined as:

Ii = Ii(x,y,z) = S(x,y,z,ti), i = 0,...,Nt-I. Dynamic enhancement curve for each pixel is calculated as: Ei = Ei(x,y,z) = Ii - Io, i = 0,...,Nt-l. In this notation E ₀ = 0. The peak enhancement for each pixel can be defined as E _max = maxi=i _j...Nt -i(Ei). The contrast arrival time (Ta) is the time required for each pixel to reach a pre-defined value of enhancement relative to Em _a x as shown in Fig. 3 for αE _max (0<α< 1).

Calculating E _max and Ta for each pixel in the current ROI, a joint distribution histogram [Jiri Jan, "Medical Image Processing, Reconstruction and Restoration: Concepts and MEthods (Signal Processing and Communications)", CRC (2005)] can be calculated for these parameters as shown in Fig. 4 (darker color represents high density of the pixels with corresponding E _max and T _a ).

Pixels with the strongest enhancement and earliest arrival times are considered to be arterial. Pixels with later arrival time (pre-defined minimum time interval between arterial and venous peaks δTav (see Fig. 4) should be set prior to processing) but still relatively strong enhancement are considered to be venous. The most enhancing part of the histogram δE ^hl is discarded, since the density of pixels in this region of the histogram is not sufficient (see Fig. 4). The resulting groups of pixels with arterial and venous enhancement patterns are indicated by circles in Fig. 4. The required number of pixels in each group can be pre-defined or made flexible by pre-setting relative enhancement and arrival time thresholds (see below). Pixels in each group are included in the order of decreasing enhancement as shown by white arrows in Fig. 4. The most enhancing pixels satisfying the conditions outlined above are included first, then peak enhancement threshold is decreased (shown by white arrows in Fig. 4) until the required number of pixels is enclosed into the group or the threshold reached its pre-defined minimum value. Averaging enhancement curves for all pixels inside each of these groups, arterial and venous reference curves R ^a j and R\ i = 0,...,Nt-I are calculated in each of the ROIs as shown in Fig. 5 based on the splitting scheme in Fig. 1. Individual steps performed by the algorithm are detailed below. 1) Finds the largest local (inside current ROI) Emax (Emaχ ^R01 )-

Then sets the initial local arterial time T^ ⁶ ^ to the corresponding value of Ta and adds this pixel to the "arterial group" which was originally empty.

2) Sets the initial enhancement threshold E _th = β*E _ma χ ^R01 (0<β<l). In Fig. 4 δE ^hi = (1- β)* E _max ^RO1 . 3) Reduces E^ by the small pre-defined small value δEth (for example δE _λ = 0.001* E _max ^R01 or δEth = 0.01 *E _th ).

4) Takes all points that have E _ma χ value in the range E _t h < E _max < (Ea ₁ + δEth).

5) Checks that Ta for all new pixels is within the pre-defined arterial time window δT ³¹ ^ (T^ - 0.5* δT ³ ^ < Ta < T ³ ^ + 0.5* δT ^aiteI3f ). If the pixel satisfies this condition, it is added to the "arterial group".

6) Recalculates the local arterial time T ^artery by averaging all Ta inside the "arterial group".

7) Checks if the total number of pixels inside "arterial group" is more or equal to the pre-defined minimum number. a) If this condition is satisfied, the generation of arterial points is stopped and step 8 starts. b) If this condition is not satisfied, then the algorithm checks whether E _t h exceeds the pre-defined minimum allowed value of E _t h (E _t h ^min = γ E _max ^R01 , 0 < γ < 1). If so, the algorithm continues to step 3. If not, the generation of arterial points is stopped and step 8 starts.

8) Averages all individual dynamic enhancement curves inside the "arterial group" to generate local arterial reference curve.

9) Arterial reference curve R ^a i is generated for current ROI. Then, algorithm removes all pixels with Ta < j ⁸ " ¹ "^ + δTav, where

δTav is the pre-defined time interval between arterial and venous arrival (shown in Fig. 4). The rest of the histogram (later enhancement part) is sent to the same algorithm as described in steps 1 - 9 (with all pre-defined parameters set to those for vein and all indices changed from arteries to veins), to generate venous reference curve R ^V j. To improve the robustness of this process, step 5 can be optionally extended to include additional check for similarity between individual pixel enhancement curve and previously generated local R ^a j. If the similarity measure (SM) (different choices for SM are defined below) exceeds pre-defined similarity threshold SM ^av _max , the pixel is discarded since it is too similar to an arterial pixel and therefore does not belong to the"vein group".

This algorithm, as described above, is fully automated and robust, since it takes into account contrast dynamic and circulation rate for each individual patient.

Generation of arterial and venous reference pixels ("seeds")

Once the reference curves for multiple ROIs are generated, a pair of correlation coefficients [B. Iordanova, D. Rosenbaum, D. Norman, M. Weiner, and C. Studholme, "MR Imaging Anatomy in Neurodegeneration: A Robust Volumetric Parcellation Method of the Frontal Lobe Gyri with Quantitative Validation in Patients with Dementia", AJNR Am. J. Neuroradiol., 27(8), pp. 1747-1754 (2006), Bock M, Schoenberg S. O., Flomer F., Schad L. R., "Separation of arteries and veins in 3D MR angiography using correlation analysis", Magn. Reson. Med. 43, pp. 481-187 (2000), Mazaheri Y, Carroll T. J., Mistretta C. A., Korosec F. R., Grist T. M., "Vessel segmentation of 3D MR angiography using time resolved acquisition curves", Proc. ISMRM 6th Scientific Meeting, Philadelphia, p. 2181 (1999)], mutual information coefficients [B. M. Klabbers et. al, "Matching PET and CT scans of the head and neck area: Development of method and validation", Medical Physics 29(10), pp. 2230-2238 (2002)], overlap coefficients [B. Iordanova, D. Rosenbaum, D. Norman, M. Weiner, and C. Studholme, "MR Imaging Anatomy in Neurodegeneration: A Robust Volumetric Parcellation Method of the Frontal Lobe Gyri with Quantitative Validation in Patients with Dementia", AJNR Am. J. Neuroradiol, 27(8), pp. 1747-1754 (2006)], or some other measurement of SM between individual enhancement curve Ei (i = 0, ... ,Nt- 1 ) and corresponding reference curves R ^a i and R ^V j (i = 0,...,Nt-I) is calculated for each pixel inside ROI. Based on these similarity parameters, any required number of reference pixels from arteries and veins can be generated. The process starts from the pixels with the highest similarity values and continues until the pre-set number of pixels is generated or other stopping criteria (including but not limited to similarity threshold, relative (absolute) local(global) enhancement- and (or) intensity threshold) is satisfied. Below is the detailed description of the steps performed by the algorithm for arterial seed generation (SMa and SMv are used below to identify SM between individual pixel enhancement curve Ej and local arterial and venous reference curves R ^a i and R ^V i).

1) Finds the pixel with highest SMa. Then sets the initial similarity measure SMath to this value SMa _max .

2) Reduces SMa _th by the small pre-defined value δSMth, so SMa _th = SMath - ASMu ₁ .

3) Takes all pixels with SMa _th < SMa < SMa ₁₁₁ + δSM _th .

4) Checks that E _max > ψ* E _n13x ¹¹⁰¹ (ψ is the pre-defined relative noise level 0<ψ <1) and (or) E _max > E _ma χ ^Nolse (E _ma χ ^No ' ^se is the pre-defined signal noise level) for all new pixels from step 3. This step discards pixels with the enhancement curve similar to R ^a j, but with non-sufficient enhancement level.

5) Checks that SMa > τ* SMv (τ > 0 is the pre-defined similarity threshold between A and V pixels). This step discards pixels that have high similarity with R ^V j.

Note: AU pixels that failed steps 4 and 5 are discarded.

6) If SMa _t h < SMa _th ^min (the pre-defined minimum SM for arteries), the algorithm stops the generation of arterial "seeds" and starts step 9.

7) (Optional step) If the total number of generated arterial "seeds" exceeds the pre-defined required number of seeds N _seeds ^a , the algorithm stops the generation of arterial "seeds" and starts step 9.

8) Starts step 2.

9) Generation of arterial "seeds" is finished.

This algorithm can be repeated for veins as described in steps 1 - 9 (with all pre-defined parameters set to those for vein and all indices changed from arteries to veins), to generate venous seeds". An example of the generated multiple reference pixels of arterial type, is shown on the right of Fig. 6 next to the original MRA image.

Segmentation of the dataset

Automatically generated arterial and (or) venous seeds can either be used for segmentation [Dellepiane, S. G. and Novelli, L. and Bruzzo, M. and Antonelli, M., "A Fuzzy Connectivity Tree for Hierarchical Extraction of Venous Structures", Scandinavian Conference on Image Analysis SCIA03, pp. 844-852 (2003), Jiri Jan, "Medical Image Processing, Reconstruction and Restoration: Concepts and MEthods (Signal Processing and Communications)", CRC (2005)] to generate arterial and (or) venous masks, or they can be used themselves as an arterial and (or) venous mask, if the number of generated seeds is enough for

smooth mapping of a vessel of a certain type. Arterial and venous masks are then stored in different color channels (for example red arteries, blue/green veins) for better 2D and 3D visualization when shown together as illustrated in Fig. 7. Storing of arterial (venous) structures in different color channels allows the user to quickly enhance (suppress) arterial (venous) components of the angiograms by changing relative channel intensities using any available image viewing/editing software.