WO2018130601A2

WO2018130601A2 - Extracting flow information from a dynamic angiography dataset

Info

Publication number: WO2018130601A2
Application number: PCT/EP2018/050629
Authority: WO
Inventors: Midas MEIJS; Frederick J. Anton MEIJER; Rashindra MANNIESING
Original assignee: Stichting Katholieke Universiteit; Technologiestichting Stw Nederlandse Organisatie Voor Wetenschappelijk Onderzoek (Nwo)
Priority date: 2017-01-12
Filing date: 2018-01-11
Publication date: 2018-07-19
Also published as: WO2018130601A3

Abstract

A system for extracting flow information from a dynamic volumetric angiography dataset is configured for segmenting (202) a vasculature structure by labeling certain voxels of the dynamic volumetric angiography dataset as belonging to the vasculature. For each of a plurality of the labeled voxels, (203) a voxel time value is calculated indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition. A histogram is calculated (204) of the calculated voxel time values. The system determines (205) a time window having a lower voxel time value and an upper voxel time value, by fitting a model to the histogram, wherein the model defines the time window in terms of certain predetermined features of the histogram. The system associates (206) a first color to the lower voxel time value, and a second color to the upper voxel time value.

Description

Extracting flow information from a dynamic angiography dataset

FIELD OF THE INVENTION

The invention relates to extracting flow information from a dynamic volumetric angiography dataset. BACKGROUND OF THE INVENTION

Stroke can be caused by a blockage or rupture of a feeding artery to the brain. With modern computed tomography (CT) scanners, 4-dimensional (4D) images can be acquired after a contrast agent injection, showing the cerebral blood flow over time.

These 4D CT images are used by neuroradiologists for diagnosis and treatment planning. Other kinds of medical imaging equipment, such as MRI, may also be used to collect such a 4D volumetric dataset. However, it is difficult for medical staff to review the massive amount of data of a dynamic volumetric angiography dataset.

"Color-Coded Cerebral Computed Tomographic Angiography", Kolja M.

Thierfelder et al., Investigative Radiology, Volume 50, Number 5, May 2015, discloses a method of displaying dynamic cerebral computed tomographic (CT) angiography (CTA) data sets in which the time delay to maximum enhancement is displayed in a range of colors. A combination of threshold application and region-growing methods was used to detect bone and nonenhancing voxels in the brain tissue. To narrow down the positive vascular mask, a reference vessel was selected with high enhancement in a central slice. An arterial input function was a subset of voxels classified as enhancing. Time attenuation curves that enhance very early and are located in an equidistant slice are denoted, and their mean constitutes the arterial input function, which is used to derive delay values that are subsequently colored and superimposed over a maximum projection image.

The large size and complexity of 4D CT images makes an accurate analysis difficult and time-consuming compared to lower dimensional imaging data. Also, there is a chance of missing abnormalities, e.g. aneurysms or vascular malformations, by novice and even trained radiologists, due to the vast amount of data and difficulties interpreting temporal information. Related to this problem is that the 4D CT images are often presented in 2D fashion (e.g. axial slices) to the medical specialist, without any pre-processing or filtering of the relevant information. SUMMARY OF THE INVENTION

An aspect of the invention is to provide an improved system to extract flow information from a dynamic volumetric angiography dataset.

To address this concern, a system for extracting flow information from a dynamic volumetric angiography dataset by a computer system is provided. The system comprises

an input unit for receiving a dynamic volumetric dataset comprising a plurality of voxels, wherein each voxel is associated with a time sequence;

an output unit for outputting colors assigned to voxels of the dynamic volumetric dataset; and

a processor configured to control:

segmenting a vasculature structure by labeling certain voxels of the dynamic volumetric angiography dataset as belonging to the vasculature;

for each of a plurality of the labeled voxels, calculating a voxel time value indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition;

calculating a histogram of the calculated voxel time values;

determining a time window having a lower voxel time value and an upper voxel time value, by fitting a model to the histogram, wherein the model defines the time window in terms of certain predetermined features of the histogram;

assigning a predetermined first color to the lower voxel time value, and a predetermined second color to the upper voxel time value;

associating a color to different voxel time values of the window by interpolating the predetermined colors; and

assigning the color associated to the voxel time value of a voxel to that voxel.

The system provides improved extraction of flow information. The

segmentation helps to preselect voxels that are involved in blood flow. Moreover, the color mapping defined by the invention extracts the most relevant portions of the flow information because of the model that defines the time window in terms of the features of the histogram of the voxel time values. The temporal information may be projected on the segmented blood vessels in an improved way and highlights any abnormalities in the vasculature defined by temporal disturbances due to e.g. occlusions, vascular malformations and/or collateral flow.

The fitting a model to the histogram may comprise determining a first voxel time value associated with a highest frequency in the histogram; determining the lower voxel time value being a predetermined first time duration earlier than the first voxel time value; and determining the upper voxel time value being a predetermined second time duration later than the first voxel time value. This model enables to extract important flow information. In certain embodiments, this model may be advantageous in particular when the number of data points is relatively large, for example larger than 10, or larger than 18.

The fitting a model to the histogram may comprise selecting N different voxel time values associated with N largest frequencies in the histogram, wherein N is an integer larger than two, determining the lower voxel time value being the smallest voxel time value among the N different voxel time values, and determining the upper voxel time value being the greatest voxel time value among the N different voxel time values; and wherein the processor is further configured to control assigning N-2 different further predetermined colors to the N-2 voxel time values among the N different voxel time values other than the smallest voxel time value and the greatest voxel time value,

wherein the color is assigned to the different voxel time values of the window by interpolating also the further predetermined colors. This model enables to extract important flow information. In certain embodiments, this model may be advantageous in particular when the number of data points is relatively small, for example smaller than 10, or smaller than 15.

The processor may be configured to control assigning a predetermined third color to a time value corresponding to a start of an acquisition of the dynamic volumetric angiography dataset; assigning a predetermined fourth color to a time value corresponding to an end of the acquisition of the dynamic volumetric angiography dataset; and assigning a color to different voxel time values inside and outside of the window by interpolating the predetermined colors. This allows to color the voxel time values inside and outside the window differently.

The voxel time value may comprises a time-to-peak value or a time-to-signal value, wherein time-to-peak is defined as a time duration from a start time to a time of a peak in the time sequence of the voxel and a time-to-signal is defined as a time duration from the start time to a first time the time sequence exceeds a certain threshold. These voxel time values have particular relevance. For example, the time- to-peak value may be particularly relevant for stroke, and the time-to-signal value may be particularly relevant for AVM, although both values may be relevant in general for both conditions.

The segmenting a vasculature structure may comprise, for each of a plurality of voxels of the dynamic volumetric angiography dataset, calculating a temporal variance of the time sequence of the voxel; calculating a threshold based on a statistic of the calculated temporal variances; using the threshold to select candidate vessel voxels; and labeling certain voxels among the candidate vessel voxels as belonging to the vasculature, based on a plurality of features of the candidate vessel voxels, the plurality of features including the temporal variance. The temporal variance was found to be particularly decisive of whether the voxel belongs to vasculature or not.

The temporal variance may be a weighted temporal variance, wherein the processor is configured to control to weight the data points of the time sequence by a weight factor that is based on an exposure associated with each data point of the time sequence. The weighted temporal variance was found to be particularly decisive of whether the voxel belongs to vasculature or not.

The weighted temporal variance may be based on an equation WTV_{x y z} =

temporal variance for a voxel (x, y, z), WTA_{x y z} denotes the weighted temporal average for a voxel (x, y, z), with WTA_X y _Z ⁼ _ιί=ο γ,_ζ,ί ' ^wi denotes the

Et

weight, with j ——— ;, wherein Cj denotes the exposure associated with a data point i, I_x y z i denotes the value of the data point i of the time sequence of the voxel

(x, y, z), and T denotes a number of data points to be evaluated.

The plurality of features of the candidate vessel voxels may further include at least one of a temporal average, a feature of an intensity histogram computed within a neighborhood of each voxel in the temporal variance image, a distance to a border of an intracranial cavity, a Hessian calculated on a weighted temporal variance data at a plurality of different scales, a certain preselected data point of the time sequence of each voxel. Such features, alone or in combination, improve the accuracy of the labeling.

According to another aspect of the invention, a method of extracting flow information from a dynamic volumetric angiography dataset is provided. The method comprises

segmenting a vasculature structure by labeling certain voxels of the dynamic volumetric angiography dataset as belonging to the vasculature; for each of a plurality of the labeled voxels, calculating a voxel time value indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition;

calculating a histogram of the calculated voxel time values;

According to another aspect of the invention, a computer program product is provided comprising instructions configured to cause a processor system to perform a method set forth herein.

The person skilled in the art will understand that the features described above may be combined in any way deemed useful. Moreover, modifications and variations described in respect of the system may likewise be applied to the method and to the computer program product, and modifications and variations described in respect of the method may likewise be applied to the system and to the computer program product.

BRIEF DESCRIPTION OF THE DRAWINGS

In the following, aspects of the invention will be elucidated by means of examples, with reference to the drawings. Throughout the drawings, the same reference numerals may be used to indicate similar items. The drawings are diagrammatic and may not be drawn to scale.

Fig. 1 shows a block diagram of a system for extracting flow information from a dynamic volumetric angiography dataset.

Fig. 2 shows a flowchart of a method of extracting flow information from a dynamic volumetric angiography dataset.

Fig. 3 shows a flowchart of a first example of fitting a model to a histogram.

Fig. 4 shows a flowchart of a second example of fitting a model to a histogram.

Fig. 5 shows a flowchart of associating additional colors with certain voxel time values.

Fig. 6 shows a flowchart of a segmentation process. Fig. 7 illustrates an example histogram of a weighted temporal variance.

Fig. 8 shows 19 acquisition times used in the example of Fig. 7.

Fig. 9 illustrates a time window and color mapping created in the example of Fig. 7 and Fig. 8.

Fig. 10 illustrates another example histogram of a weighted temporal variance. Fig. 1 1 illustrates an example of a color-map for the example of Fig. 10.

DETAILED DESCRIPTION OF EMBODIMENTS

Certain exemplary embodiments will be described in greater detail, with reference to the accompanying drawings.

The matters disclosed in the description, such as detailed construction and elements, are provided to assist in a comprehensive understanding of the exemplary embodiments. Accordingly, it is apparent that the exemplary embodiments can be carried out without those specifically defined matters, also, well-known operations or structures are not described in detail, since they would obscure the description with unnecessary detail.

According to an aspect, underlying vessel segmentation is obtained by local histogram analysis of the temporal variance image to ensure that small vessels distally in the tree are included in the segmentation.

According to another aspect, the color mapping is centralized on the modus of the time to peak histogram to achieve patient specific normalization of the color scale, such that any deviations due to temporal disturbance are immediately visible in the visualization.

According to another aspect, a new visualization technique is provided by mapping temporal information on the blood vessels and normalizing the color scale within the patient. This facilitates the detection of pathologies including, but not limited to, occlusions, artery-venous malformations and collateral flows. The technique may comprise the following steps: first, segmentation of the cerebral vasculature from 4D CT images, second, Gaussian filtering in temporal direction to reduce image noise, third, constructing the time-to-peak (TTP) histogram of the segmented vessel voxels, and fourth, using the modus of the TTP histogram to centralize the color mapping.

According to another aspect, a color map may be defined that associates each TTP value with a particular color, so that the dynamics may be properly visualized. In a particular example implementation, the color map may be defined on a fixed time interval, for example from 0 to 60 seconds counted from the start of a dynamic acquisition. This start time may be selected, for example based on the time of contrast agent injection. Fixed colors may be assigned to these values of 0 and 60 seconds. For example, the color blue may be assigned to 0 seconds, and the color purple may be assigned to 60 seconds. The color of two more points is imposed in the spectrum with a fixed color, according to the generated spectrum. The first point, fixed on for example the color green, may be imposed approximately 6 seconds (which corresponds, in certain implementations, to 3 acquisitions) before the time of the modus of the histogram. The second point, fixed on for example the color red, may be imposed approximately 2 seconds (which corresponds, in certain implementations, to 1 acquisition) after the modus of the histogram. Moreover, the value of the arrival time may be calculated with the available time intervals between the acquisitions.

In certain implementations, after the vessels have been segmented, a coloring may be applied as follows, starting with a dataset of voxel time values of the voxels that have been labeled as vessel voxels. An optional temporal Gaussian filtering may be performed on the voxel time values to reduce temporal image noise. Then, a histogram of the voxel time values may be computed. The N largest bins of the histogram may be determined, wherein N is a predetermined positive integer, preferably an integer greater than 1 , more preferably greater than 2. N predetermined fixed colors are provided in a given order, for example from a look-up table. Each of the N predetermined fixed colors is assigned to a different one of the N largest bins, in order of the voxel time value associated with each bin.

In certain implementations, after the vessels have been segmented, a coloring may be applied as follows.

Temporal Gaussian filtering (e.g. σ = 2 acquisition samples) may be performed on the voxel time values to reduce temporal image noise.

For patients, for example with a suspicion of stroke, who received a 4D CT acquisition containing for example 19 (or more) time points, the time to peak (TTP) may be calculated for every voxel in the image. The TTP image may be masked with vessels and a histogram may be calculated with bin size of for example 1 second. Thus only the TTP values of vessel voxels are included in the histogram.

For patients with, for example, a suspicion of artery-venous malformations who received a 4D CT acquisition containing, for example, 14 (or less) time points, the time to signal (TTS) may be calculated, which is the first time that the intensity exceeds a voxel specific threshold, which is defined as T = I_mi_n + (SJmax ~ ^min) ' 0.2), with l_max and l_mi_n maximum and minimum intensity value over the time curve per voxel. The TTS image may be masked with vessels and the histogram may be calculated with a bin size of 1 second. Thus only the TTS values of vessel voxels are included in the histogram.

In both cases, a color mapping may be created as follows. Select the N largest

(non-empty) bins of the histogram. A suitable value of N may be 6. However, N can be any positive integer. If there are less than N bins then select those bins. Assign colors from bin left (corresponding to the smallest voxel time value (e.g. TTP or TTS) of the N voxel time values) to bin right (corresponding to the greatest voxel time value (e.g.

TTP or TTS) of the N voxel time values). Suitable colors may be magenta, red, yellow, green, cyan, blue (#FF00FF, (#FF0000, #FFFF00, #00FF00, #00FFFF, #0000FF). The colors in between the fixed points may be interpolated according to the Hue Saturation

Lightness (HSL) cylindrical-coordinate system.

Fig. 1 illustrates an example implementation of a system for extracting flow information from a dynamic volumetric angiography dataset by a computer system.

The system may be a computer system for example. The system may be embodied in a scanning device, that is capable of performing dynamic angiographic volumetric measurements. An example of such a device is a such as a CT scanning device. For example the system of Fig. 1 may be a subsystem of such a scanning device.

Alternatively, the system may be a standalone system that is capable of receiving a dynamic volumetric angiographic dataset that has been acquired by means of such a scanning device. The system may comprise an input unit 101 , an output unit 102, a processor 103, and a memory 104.

The input unit 101 may be any input unit capable of receiving a dynamic volumetric dataset comprising a plurality of voxels, wherein each voxel is associated with a time sequence. For example, the input unit may comprise a communications port to receive the dataset via a network connection or a wired or wireless

communication connection to a storage server or a scanner. The input unit 101 may also comprise a reading device capable to read the dataset from a removable media.

The input unit may, alternatively or additionally, comprise the scanning device. The memory 104 may optionally be configured to store temporary information, such as the dynamic volumetric angiographic dataset 105 received via the input unit 101 , and/or data derived from that dataset 105.

The output unit 102 may comprise any output unit capable of outputting colors assigned to voxels of the dynamic volumetric dataset. These colors may be outputted in form of data values representing the colors, for example. Alternatively, the colors may be outputted in form of actually rendered colors using, for example, a display device such as an LCD or an OLED screen. For example, the output unit may comprise a communications port to transmit the color values via a network connection or a wired or wireless comunication connection to a storage server or workstation. The output unit 102 may also womprise a writing device capable to store the output values on a removable media. The output unit may, alternatively or additionally, comprise a display or a printer to output the colors in a rendered form.

The processor 103 may comprise any suitable processor capable to control processing operations. For example, a microprocessor or a controller may be used. An example of such a microprocessor is an Intel Core i7 processor, manufactured by Intel Corporation. However, any suitable processor may be used. The processor 103 may comprise a plurality of microprocessors configured to cooperate with each other to perform processing operations together. The memory 104 may comprise a computer program code 107 stored thereon, which computer program code causes the processor to perform certain operations. This way, a realization of a method disclosed herein may be realized by means of such computer program code 107.

The memory 104 may comprise a volatile and/or a non-volatile memory, such as RAM, ROM, FLASH, magnetic disk, optical disk, or a combination thereof. The program code 107 may typically be stored on a non-transitory computer readable media.

The processor may be configured to control the operations of the input unit 101 , output unit 102, and memory 104. Moreover, other components may be controlled (not shown). Also, a scanning device or a display controller (not shown) may be controlled by the processor 103 to generate the inputted datasets and render the outputted datasets.

Fig. 2 shows a flowchart of a method of extracting flow information from a dynamic volumetric angiography dataset. As mentioned above, the method may be embodied as a computer program code 107, although other implementations of the method are equally possible.

In step 201 , a dynamic volumetric angiography dataset is received. For example, this step is implemented by causing the processor to control the input unit 101 to receive the dataset. For example, the dynamic volumetric angiography dataset comprises a dataset that shows a contrast agent in the vasculature as from the time it is injected up to when it has been flushed out by blood flow, for example. For each voxel, the volumetric dataset may comprise a time series or time sequence of data points. Each data point may correspond to the voxel value in one of a series of subsequent three-dimensional datasets (e.g. CT datasets). Thus, a time dependent behavior of the voxel is represented by the time sequence. In step 202, a vasculature structure represented by the dataset is detected. In this process, certain voxels of the dynamic volumetric angiography dataset are labeled as belonging to the vasculature, in dependence on the data points that have been measured for each voxel. An example of how to detect the vasculature is described elsewhere in this disclosure.

In step 203, for each of a plurality of the labeled voxels, a voxel time value is calculated. A voxel time value is indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition. Thus, it is a quantity that indicates a specific time at which the time sequence exhibits a particular feature. Specific examples of voxel time values are time-to-peak and time-to-signal. Time-to-peak may be defined as a time duration from a start time to a time of a peak in the time sequence of the voxel. Time-to-signal may be defined as a time duration from the start time to a first time the time sequence exceeds a certain threshold. Other voxel time values may be envisioned.

In an example, the threshold used to determine the time-to-signal value of a particular voxel may

wherein I_mi_n denotes the smallest intensity value of a data point of a time series of the voxel, and I_max denotes the largest intensity value of a data point of a time series of the voxel. It will be understood that in the above equation, the factor 0.2 may be replaced by any suitable value between 0 and 1 .

In step 204, a histogram of the calculated voxel time values is calculated. Thus, the voxel time values of the different voxels are collected and the distribution thereof is computed in form of a histogram.

In step 205, a time window is determined by fitting a model to the histogram.

The time window typically may represent the time interval in which most relevant information is to be found. Therefore, the information in this time window may be outputted in greater detail. The model defines the time window in terms of certain predetermined features of the histogram known to be related to clinically relevant information. The time window has a lower voxel time value and an upper voxel time.

In step 206, a predetermined first color is associated to the lower voxel time value, and a predetermined second color is associated to the upper voxel time value.

The predetermined first color is different from the predetermined second color.

Preferably, the two colors are capable of being easily distinguished by a human observer. In a particular example, green (hexadecimal RGB code 00FF00) is associated with the lower voxel time value and red (hexadecimal RGB code FF0000) is associated with the upper voxel time value.

In step 207, the other voxel time values (to which any predetermined color has not yet been associated), are associated with colors by interpolating the

predetermined colors. Thus, suppose a first voxel time value v₁ has been associated with a color with RGB code (T , , ¾i) , and a second voxel time value v₂ has been associated with a color with RGB code (r₂, g₂, b₂ Then, a third color v₃ that is larger than v₁ and smaller than v₂ can be expressed as v₃ = xv^ + (1 — x)v , for a certain value x, wherein 0 < x≤ 1. In that case the color v₃ may be associated with the color with RGB code (xr_x + (1 — x)r₂, xg₁ + (1 — x)g₂, xb + (1— x)fc₂)- Other kinds of interpolation are also possible. For example, the colors in between the fixed points may be interpolated according to a Hue Saturation Lightness (HSL) cylindrical-coordinate system.

For example, the voxels in the window may be interpolated based on the lower voxel time value and the upper voxel time value. When more predetermined colors have been associated with particular voxel time values, these predetermined colors may be involved in the interpolation.

In step 208, the color associated to the voxel time value of a voxel may be assigned to that voxel. That is, for each voxel the voxel time value is evaluated and the associated color is looked up or calculated. Then, that color is assigned to that voxel. This process may be performed to all voxels labeled as being part of the vasculature.

In step 209, the colors assigned to the voxels of the dynamic volumetric dataset may be output via the output unit 102. For example, the colors may be exported as volumetric dataset in which the color information is attached to each voxel. Visual renderings of the volumetric dataset may be created and displayed, wherein the colors of assigned to the voxels may be applied. The colors may be merged with or superimposed on a maximum intensity projection (MIP) of a volumetric dataset representing the structural aspect of the vasculature. Rendering methods involving colored volumetric datasets are known to the person skilled in the art per se and do not form part of the present disclosure.

Fig. 3 illustrates a first example of step 205, fitting a model to the histogram. In step 301 , a first voxel time value is determined. This first voxel time value may be described as the mode (i.e. the largest peak) of the histogram with corresponding voxel time value f. In step 302, the lower voxel time value t_jower indicating the lower bound of the time window is set. This value is set as the voxel time value f minus a predetermined quantity, i.e., £_jower = £— At_{l t} for a certain predetermined positive value At^. Similarly, in step 303, the upper voxel time value t_upper indicating the upper bound of the time window is set. This value is set as the voxel time value f plus a predetermined quantity, i.e., ., t_upper = £ + At₂, for a certain predetermined positive value At₂. In a practical implementation example, At^ may be about three times larger than At₂. This may be the case, for example, when the application is stroke detection. In a specific example protocol for stroke, At^ may be about 4 seconds, and At may be about 2 seconds.

Fig. 4 illustrates a second implementation example of steps 205, fitting a model to the histogram, and step 206, associating predetermined colors to specific voxel time values. This example involves a parameter N, which represents the number of bins of the histogram that should be involved in the window. In this example, the number N is at least 3. It will be understood that a histogram bin may equivalently be expressed as a number of voxels in the bin, i.e., a frequency, or as the fraction of the total number of voxels that are in the bin, i.e., a relative frequency. In step 401 , the N bins with the largest frequencies are selected. Each bin is associated with a particular range of voxel time values. For each bin, a representative voxel time value may be chosen (for example, the center voxel time value of the bin). In step 403, the lower voxel time value is determined. To this end, the smallest voxel time value among the representative voxel time values of the N selected bins is set to be the lower voxel time value.

In step 404, the upper voxel time value is determined. To this end, the largest voxel time value among the representative voxel time values of the N selected bins is set to be the upper voxel time value.

In step 405, it is illustrated that not only the upper voxel time value and the lower voxel time value are associated with predetermined colors, but also the representative voxel time values of the remaining N-2 selected bins are assigned different predetermined colors. This is done in order. For example, if the voxel time values, in ascending order, are v_lt ... , v_N and the predetermined colors are

c_lt ... , c_N , then the color c₁ is associated with the voxel time value v₁ , the color c₂ is associated with the voxel time value v₂, and so forth until the color c_N, which is associated with voxel time value v_N. Herein, v₁ is the lower voxel time value and v_N is the upper time value. Each of the colors c₁, ... , c_N are different. As mentioned above, colors associated with other voxel time values may be obtained by

interpolation.

Fig. 5 illustrates additional steps that may optionally be performed for the predetermined colors in step 206. In addition to the colors based on the model of the histogram (e.g. the ones disclosed in Fig. 3 or 4), at least two additional colors may be associated with certain voxel time values. That is, in step 501 , a predetermined third color is associated to a time value corresponding to a start of an acquisition of the dynamic volumetric angiography dataset. This predetermined third color can be a different color than the other predetermined colors, or it can be the same color as the color associated with the lower voxel time value. The start of the acquisition may be, for example, the time of the first data point in the time sequence of a voxel.

Alternatively, it may be the time of a contrast agent injection, for example. In step 502, a predetermined fourth color is associated to a time value corresponding to an end of the acquisition of the dynamic volumetric angiography dataset. This predetermined fourth color can be a different color than the other predetermined colors, or it can be the same color as the color associated with the upper voxel time value. The end of the acquisition may be, for example, the time of the last data point in the time sequence of a voxel. Alternatively, it may be a predetermined amount of time after a contrast agent injection, for example. It will be understood that, in step 207, color values can be associated to voxel time values for voxel time values inside and outside of the window by interpolating the appropriate predetermined colors.

Fig. 6 illustrates an example implementation of the segmentation process of step 202. The segmentation process is performed on a dataset containing, for example, Hounsfield units of a scanned object. Alternatively, other kinds of scanned data such as MRI data may be segmented. Since the dataset captures dynamic information, there may be motion disturbances. To remove the effect of motion, before the actual segmentation, optionally an image alignment may be performed (step 601 ). For example, all images may be aligned with the first image. The alignment may be a rigid alignment. Alternatively, non-rigid alignment may be used. An example of a suitable registration algorithm is disclosed in Manniesing, R., Leemput, S. van de, Prokop, M. and Ginneken, B. van 2016. White Matter and Gray Matter Segmentation in 4D CT Images of Acute Ischemic Stroke Patients: a Feasibility Study. Annual Meeting of the Radiological Society of North America (2016).

After the optional alignment, a temporal variance may be computed for each voxel, in step 602. This may be a weighted temporal variance. The weighted temporal variance may be weighted with weights that depend on the exposure. In a non- weighted embodiment, the weights in the below formulas may be all equal, i.e. _έ = 1, or wi = 1/T, wherein T is the last data point. Weights depending on the exposure may be, for example, defined as follows:

Other weight values may be computed and used similarly, for example based on instead of Ej . Herein, i may range from 1 to T.

In an example implementation, the weighted temporal variance of a voxel may be computed in two steps. First a weighted temporal average may be computed, for example as follows:

T

WTA x,y,z ^wi ^' Ιχ,γ,ζ,ί·

Herein, "x, y, z" denotes a voxel with coordinates (x, y, z). The symbol i denotes again the sample point, and T denotes the last data point. l_Xiy_iZ,i denotes the data point i of the voxel (x, y, z).

Second, the weighted temporal variance may be computed, for example, as follows:

WTV_Xiy_iZ =

In step 603, a threshold is calculated based on the histogram of the weighted temporal variance image. This threshold determines which voxels will be considered as candidate vessel voxels. The threshold may be a predetermined value or may be based on the histogram of the weighted temporal variance image. For example, the threshold may be based on a statistic of the voxels of the weighted temporal variance image. In a particular example, the threshold is based on the mean and the standard deviation, for example as follows:

threshold = mean of WTV + C x standard deviation of WTV, wherein C is a constant, for example 1.5, and WTV is the weighted temporal variance image. In step 604, the voxels having a weighted temporal variance greater than the threshold, may be selected as candidate vessel voxels. In step 605, several features of the dynamic volumetric angiography dataset are determined, for different voxels of the dataset. The features may be determined for only the candidate vessel voxels. In another implementation, the features may be determined for all voxels. Examples of relevant features are:

• Weighted temporal average:

Weighted temporal variance:

Local histogram features: An intensity histogram may be computed within a neighborhood around each voxel in the WTV image. The mean, standard deviation, modus, and entropy may be computed from these histograms, and they can be used as features. The features may be computed for

neighborhoods of different sizes. For example, the features may be computed for a 5x5x5 and 9x9x9 neighborhood around a certain voxel.

Distance to the border of the intracranial cavity: For example, the Euclidean distance to the border of the intracranial cavity may be computed for each voxel. The border of the intracranial cavity may be calculated by means of a suitable segmentation method that finds the border between the cranium and grey/white matter. Such a method is known in the art by itself. Distances from inside the intracranial cavity to the border may be denoted as negative distances, and distances from outside the intracranial cavity to the border may be denoted as positive distances. In an alternative implementation, the sign may be inverted.

Hessian: The eigenvalues of the Hessian system may be calculated on the WTV at a plurality of different scales. For example, four scales may be selected on an evenly distributed range from 0.5 to 2.0 mm.

TO: The first time point volume, which comprises the intensity of the first data point of the time sequence of each voxel, may be used to represent the tissue with minimal contrast. Alternatively, another predetermined data point may be selected instead of the first data point. In step 606, the candidate vessel voxels are classified as vessel voxel or non- vessel voxel based on the determined features. This can be performed by means of different classification techniques, such as a neural network, nearest neighbor, or a random forest classifier. Random forest classifier is known from, for example, Breiman, L, 2001. Random forests. Machine learning 45 (1 ), 5-32. For example, a random forest classifier with 100 tree and maximum tree depth of 30 may be used. The classifier may be trained with a training dataset, as is known to the person skilled in the art.

In step 607, optional post-processing steps may be performed, such as a connected component analysis (for example, based on a 26-neighborhood), while discarding components smaller than 25 voxels. Alternatively, another size

neighborhood and minimum number of voxels per component may be used. Another optional post-processing step is morphological hole filling. For example, a 3x3x3 kernel size is used, with a plurality of for example 10 iterations.

Fig. 7 illustrates an example histogram of a voxel time values calculated in step

204. In this particular example the voxel time values are time-to-peak values. A bin size of one second is used, whereas the interval between successive acquisitions is larger than one second. Therefore some bins are empty in this particular case. The horizontal axis represents the representative time-to-peak associated with a bin and the vertical axis represents the number of voxels in the bin. The largest bin, or modus, of this histogram is found for a time-to-peak of 28 seconds.

Fig. 8 shows the 19 acquisition times used in the example of Fig. 7. The horizontal axis represents acquisition time, and the dots 801 indicate the acquisition times.

Fig. 9 illustrates a time window and color mapping created in steps 206, 207,

301 , 302, 303 in the example of Fig. 7 and Fig. 8, based on the TTP histogram. The acquisition corresponding to the modus of the histogram is the 13^th. The time window is then defined with the start of the time window at the 13-3 = 10^th acquisition (at 21 .8 seconds = green), and the end of the time window is at 13+1 = 14^th acquisition (at 30.3 seconds = red). The first acquisition at 1 .0 seconds is set to green and the last acquisition at 54.3 seconds is set to purple, in accordance with steps 501 , 502.

Fig. 10 illustrates another example histogram of a weighted temporal variance, in this case of a time-to-signal calculated in step 204. The time-to-signal is based on a threshold T_signal = l_min + ((I_max - I_min) ^■ 0.2). The number of acquisitions is 14. The bin size of the histogram is 1 second. The 6 largest bins are determined at 13.5, 14.5, 12.5, 15.5, 16.5 and 17.5 seconds. The horizontal axis reprsents the representative time-to-signal of the bin, the vertical axis the number of voxels in the bin.

Fig. 1 1 illustrates an example of a color-map for the example of Fig. 10, calculated in accordance with steps 206, 207, 401 , 403, 404. This example results in the following color-map: magenta at 12.5s, red at 13.5s, yellow at 14.5s, green at

15.5s, cyan at 16.5s and blue at 17.5s. The times of the first and last acquisitions have been set to zero (black), in accordance with steps 501 , 502.

It will be noted that both the method of Fig. 3 and the method of Fig. 4 can be applied to any kind of voxel time value or number of acquisitions. For example, the method of Fig. 4 can be applied to both time-to-peak values and time-to-signal values, and can be used for all kinds of applications such as stroke and artery-venous malformations.

Some or all aspects of the invention may be suitable for being implemented in form of software, in particular a computer program product. The computer program product may comprise a computer program stored on a non-transitory computer- readable media. Also, the computer program may be represented by a signal, such as an optic signal or an electro-magnetic signal, carried by a transmission medium such as an optic fiber cable or the air. The computer program may partly or entirely have the form of source code, object code, or pseudo code, suitable for being executed by a computer system. For example, the code may be executable by one or more processors.

The examples and embodiments described herein serve to illustrate rather than limit the invention. The person skilled in the art will be able to design alternative embodiments without departing from the spirit and scope of the present disclosure, as defined by the appended claims and their equivalents. Reference signs placed in parentheses in the claims shall not be interpreted to limit the scope of the claims. Items described as separate entities in the claims or the description may be implemented as a single hardware or software item combining the features of the items described.

Claims

CLAIMS:

1 . A system for extracting flow information from a dynamic volumetric

angiography dataset, comprising

an input unit (101 ) for receiving a dynamic volumetric dataset comprising a plurality of voxels, wherein each voxel is associated with a time sequence;

an output unit (102) for outputting colors assigned to voxels of the dynamic volumetric dataset; and

a processor (103) configured to control:

segmenting (202) a vasculature structure by labeling certain voxels of the dynamic volumetric angiography dataset as belonging to the vasculature;

for each of a plurality of the labeled voxels, calculating (203) a voxel time value indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition;

calculating (204) a histogram of the calculated voxel time values;

determining (205) a time window having a lower voxel time value and an upper voxel time value, by fitting a model to the histogram, wherein the model defines the time window in terms of certain predetermined features of the histogram;

associating (206) a predetermined first color to the lower voxel time value, and a predetermined second color to the upper voxel time value;

associating (207) a color to different voxel time values of the window by interpolating the predetermined colors; and

assigning (208) the color associated to the voxel time value of a voxel to that voxel.

2. The system of claim 1 , wherein the fitting a model to the histogram (205) comprises:

determining (301 ) a first voxel time value associated with a highest frequency in the histogram;

determining (302) the lower voxel time value being a predetermined first time duration earlier than the first voxel time value; and

determining (303) the upper voxel time value being a predetermined second time duration later than the first voxel time value.

3. The system of claim 1 , wherein the fitting a model to the histogram (205) comprises:

selecting (401 ) N different voxel time values associated with N largest frequencies in the histogram, wherein N is an integer larger than two,

determining (403) the lower voxel time value being the smallest voxel time value among the N different voxel time values, and

determining (404) the upper voxel time value being the greatest voxel time value among the N different voxel time values; and

wherein the processor is further configured to control associating (405) N-2 different further predetermined colors to the N-2 voxel time values among the N different voxel time values other than the smallest voxel time value and the greatest voxel time value,

wherein the color is associated (207) to the different voxel time values of the window by interpolating also the further predetermined colors.

4. The system of claim 1 , 2, or 3, wherein the processor is further configured to control

associating (501 ) a predetermined third color to a time value corresponding to a start of an acquisition of the dynamic volumetric angiography dataset;

associating (502) a predetermined fourth color to a time value corresponding to an end of the acquisition of the dynamic volumetric angiography dataset; and

associating (207) a color to different voxel time values inside and outside of the window by interpolating the predetermined colors.

5. The system of claim 1 , wherein the voxel time value comprises a time- to-peak value or a time-to-signal value, wherein time-to-peak is defined as a time duration from a start time to a time of a peak in the time sequence of the voxel and a time-to-signal is defined as a time duration from the start time to a first time the time sequence exceeds a certain threshold.

6. The system of claim 1 , wherein the segmenting a vasculature structure comprises:

for each of a plurality of voxels of the dynamic volumetric angiography dataset, calculating (602) a temporal variance of the time sequence of the voxel;

calculating (603) a threshold based on a statistic of the calculated temporal variances;

using the threshold to select (604) candidate vessel voxels; and labeling (606) certain voxels among the candidate vessel voxels as belonging to the vasculature, based on a plurality of features of the candidate vessel voxels, the plurality of features including the temporal variance.

7. The system of claim 6, wherein the temporal variance is a weighted temporal variance, wherein the processor is configured to control to weight the data points of the time sequence by a weight factor that is based on an exposure associated with each data point of the time sequence.

8. The system of claim 7, wherein the weighted temporal variance is

WTV_{x y z} denotes the weighted temporal variance for a voxel (x, y, z),

WTA _{y z} denotes the weighted temporal average for a voxel (x, y, z), with

Wj denotes the weight, with Wj ——— ;, wherein E_i denotes the exposure associated with a data point i,

Ιχ,γ,ζ,ί denotes the value of the data point i of the time sequence of the voxel (x, y, z), and

T denotes a number of data points to be evaluated.

9. The system of claim 7, wherein the plurality of features of the candidate vessel voxels further includes at least one of a temporal average, a feature of an intensity histogram computed within a neighborhood of each voxel in the temporal variance image, a distance to a border of an intracranial cavity, a Hessian calculated on a weighted temporal variance data at a plurality of different scales, a certain preselected data point of the time sequence of each voxel.

10. A method of extracting flow information from a dynamic volumetric angiography dataset by a computer system, comprising

segmenting (202) a vasculature structure by labeling certain voxels of the dynamic volumetric angiography dataset as belonging to the vasculature; for each of a plurality of the labeled voxels, calculating (203) a voxel time value indicative of a time when a time sequence of a voxel of the dynamic volumetric angiography dataset satisfies a certain predetermined condition;

calculating (204) a histogram of the calculated voxel time values;

1 1 . A computer program product comprising instructions configured to cause a processor system to perform the method of claim 10.