SE516851C2 - Processing method for ventilation perfusion lung scintigrams, using image intensities to derive contours for generating set of adapted images - Google Patents

Abstract

A set of adapted digital ventilation and perfusion images is generated by adapting the images to each other based on contours (11) in these images. These contours are derived from intensities in the images. Independent claims are also included for (a) a memory medium used to carry out this method, which can be read by a computer or stored on a computer program, and (b) an image processing device for carrying out this method.

Description

The perfusion scintigram, also called the perfusion image, is generated by giving a patient an intravenous injection of, for example, a Tc-99m radionucleotide-labeled human serum albumin (see Fig. 1).

Usually these scintigrams are then analyzed manually. A normal pulmonary scintigram, ie when the ventilation and perfusion picture are in good agreement, excludes pulmonary embolism. A pulmonary scintigram in pulmonary embolism, ie for a person with pulmonary embolism, shows differences between the ventilation and perfusion images. These differences arise because the pulmonary embolism prevents the blood from reaching all parts of the lung and the perfusion image then does not cover the same surface as the ventilation image. The perfusion image will thus have unexposed, or poorly exposed, areas in areas in which a pulmonary embolism is present. These unexposed areas will not be present in the ventilation image and there is a difference between the images. If the pulmonary embolism is small, or is located at the edge of the lung, it may be difficult to distinguish these darker parts from a lung with atypical anatomy (see Figs. Lb and ld). For a more detailed description of the method, see "Diagnosis of Pulmonary Embolism", Wollmer P., Appl.

Cardiopumonary Patophysiology 2: 13-22, 1998.

A problem with lung scintigraphy is that the patient must exhale and clear the lungs of the radioactive material the patient has previously inhaled before new radioactive material is injected into the bloodstream. Due to this, the patient will often not assume exactly the same position in front of the camera and the ventilation image is then taken from a different angle than the perfusion image. A direct comparison »between the images, such as a direct subtraction of them, then does not become fair because it is not the same parts of the lung that are compared.

Another problem is the naturally occurring differences in the lungs, especially in the lung edges, which differences make the comparison difficult. Performing this diagnosis manually with great reliability requires a great deal of experience. Since doctors are often completely left to their intuition, different doctors often assess the X-rays in different ways, and also differently from time to time.

In addition, doctors often use different criteria in assessments, which makes it difficult to build up a common experience base in healthcare. There is thus a need for an automated and standardized way of diagnosing pulmonary embolism and also other lung diseases.

An automated procedure is described in "Automated interpretation of ventilation-perfusion lung scintigrams for the diagnosis of pulmonary embolism using artificial neural networks", Holger Holst et al, European Journal of Nuclear Medicine, vol. 27, no. 4, April 2000. This for The procedure is to first generate a ventilation image and then a perfusion image.The pixels in the images that have an intensity that exceeds a predetermined threshold value are removed and replaced with a lower value.Then the images are transformed in their entirety - heat, ie all image material in the image, according to finished lung templates that represent "normal" lungs. Then a quota image is generated between the perfusion and ventilation image. Each pixel in this quota image is the ratio of the corresponding pixels in the perfusion and ventilation images. - it is divided into 18 different segments and the quotas within these segments are added together into subtotals, one for each segment, and these subtotals are divided see day with the number of pixels in each segment. If this ratio is too small, this indicates that there is a possible pulmonary embolism. However, this method is unstable, ie its result is sensitive to variations in input data and thus does not work satisfactorily when the substrate is not of the highest quality, which is often the case, and not reliable enough to be used in real situations; It is also a slow process.

Today's procedures are therefore not suitable for automation with computers as the procedures leave behind many errors, such as those described above. Thus, there is a need for a method and apparatus that are simple to automate and provide reliable results quickly.

SUMMARY OF THE INVENTION An object of this invention is to provide a method for processing pulmonary scintigraphy images, which is suitable for automation.

Another object of this invention is to provide a method which enables rapid and reliable detection of pulmonary embolism.

A further object of this invention is to provide an apparatus and a memory medium having a program for performing these methods.

These objects are fulfilled by a method, a device and a memory medium according to the appended claims.

According to one aspect of the invention, there is provided a method of treating a digital ventilation image showing the airflow in a lung and a digital perfusion image showing the blood flow in the lung. The method is characterized in that a set of adapted images is generated by the images being adapted to each other based on contours in the images, which contours are based on the intensity of the images.

It will be appreciated that the custom images may be either new images or the original images in which changes have been made.

This gives the advantage that a smaller amount of data is used in the adaptation, which provides both faster and more secure adaptation. The increased safety is partly due to the fact that it is a smaller image material that must be adapted and consequently there are thus fewer sources of error that can be caused by a pulmonary embolism.

Preferably, the contours are generated by means of a threshold, whereby this contour is used in the adjustment.

It should be understood that other methods of generating the contours, such as edge detection, can also be used. ønmoo ~ v »~ lO 15 20 25 30 35 5-16 .851 5 Alternatively, the contours are shown in the pictures before starting the treatment.

The adjustment of the lung contour is advantageously done with a temporary image content that depends on a function of the contour. Advantageously, a function is used which has a constant value, preferably the value 1 inside the contours and the value 0 outside. Such a function facilitates and improves the adaptation. It is of course possible to use other functions as well.

The steps described above provide a method that is well suited for automation and provides reliable results.

Preferably, an intensity scaling is performed on one of the images to compensate for defects, which are the result of, for example, poor inhalation or for short recording times, by calculating an intensity quantification in the range 50-95% and then re-scaling the intensities of the ventilation and perfusion image so that these quantiles become equal.

This step is based on the fact that the intensity in areas with defects is below this quantile, which is why the intensity scaling takes place without regard to any defects.

This gives the advantage that it is easier to see the difference between images where a lot of air and little blood has flowed and images where little air and no blood has flowed (as in obstructive lung disease, because it is then recognized that it is difficult to find pulmonary emboli due to difficult to interpret images).

In the ventilation images, there are sometimes so-called “hot spots”, ie areas with high intensity due to the radioactive aerosol being stuck in the same place. These “hot-spots” result in large differences between the ventilation image and the perfusion image, which may be misinterpreted as emboli.

Preferably, the intensity in these "hot spots" is reduced or replaced, i.e. pixels in an image which have an intensity higher than a 90-99% intensity quantile.The intensity in these points is reduced or replaced by a ...- 10 15 20 25 30 35 ~ aau ou 5 1,6 8,51 6 lower value, eg by interpolation with, for example, bicubic or Laplace interpolation.

On the one hand, this gives the advantage that the intensity scaling in the step above becomes smoother and closer to the intensity median, which provides a more stable procedure, and on the other hand, it reduces the risk that so-called “hot spots” give rise to misinterpretations.

The large difference between a “hot spot” in one image and a pixel with normal intensity in the other image can easily be misinterpreted in a comparison between the images as meaning that the “hot spot” is a pixel with normal intensity and the pixel with normal intensity has a very low intensity, which would indicate an embolism.

Before the adjustment is performed, it is advantageous to smooth out the contour of a lung image by removing any protrusions in the image that correspond to radioactivity from the neck or stomach. These committees or contributions otherwise make it difficult to properly adjust the lungs, as these contributions are often missing in the perfusion picture. These contributions are connected to the lung via narrower segments and can thus be removed if segments that are narrower than a predefined value are excluded.

Advantageously, the contour of an image is replaced by its convex envelope before the adjustment is performed.

This step has the advantage that pulmonary emboli on the edge, which give rise to wedge-shaped defects in the perfusion image, do not affect the adaptation of the lungs.

These steps can all be performed individually or in arbitrary combinations in many different ways and to some extent in different order.

According to another aspect of the present invention, there is provided a computer program which is stored on a memory medium which can be read by a computer. The computer program is intended to process a digital ventilation image, which shows the air flow in a lung, and a digital perfusion image, which shows the blood flow in the lung. The program is characterized in that it is arranged to generate a set of adapted images by adapting the images to each other based on contours in the images, which contours are based on the intensity of the images.

It should be understood that this computer program provides the same benefits as the procedure above and can be adapted to the procedure described above.

According to another aspect of the present invention, there is provided an image processing apparatus for treating a lung image, the apparatus having means for inputting a digital ventilation image showing the air flow in a lung, and a digital perfusion image showing the blood flow in a lung. The device is characterized in that the device is arranged to generate a set of adapted images by the images being adapted to each other based on the contours of the images, which are based on the intensity of the images.

This image processing device can advantageously be implemented with a computer.

Preferably, the custom images as above are used in conjunction with an artificial neural network for automatic diagnosis.

It will be appreciated that this image processing device provides the same advantages as the above method and can be adapted to the method described above.

Brief description of the drawings The invention will be further described in the following by means of exemplary embodiments with reference to the accompanying drawing.

Fig. 1 are four radiograms showing a ventilation image (1a) and a perfusion image (1b) of a healthy lung, as well as a ventilation image (1c) and a perfusion image (1d) of a lung in which there is pulmonary embolism.

Fig. 2 is a flow chart showing an alternative embodiment of this invention.

Fig. 3 is a diagram showing a contour and its convex casing.

Fig. 4 is a series of diagrams showing adaptations of two image pairs and their covariance. -vo ~> - -. -. -uns ~ 10 15 20 25 30 35 516 8.51 8 Fig. 5 is a diagram showing a contour with neck braces and the same contour with neck braces removed.

Fig. 6 is a flow chart showing an alternative embodiment of this invention.

Fig. 7 is a flow chart showing an alternative embodiment of this invention.

Fig. 8 is a flow chart showing a preferred embodiment of this invention.

Fig. 9 is a block diagram schematically showing a system according to an alternative embodiment of this invention.

Detailed Description of the Invention The present invention will now be described, by way of example, with reference to the accompanying drawings.

An embodiment of the invention is shown schematically in Fig. 2. Based on the contour 1 of an image (see Fig. 3), which is based on, for example, the intensity of the image, a convex envelope 2 is formed in step 201. A convex envelope is the smallest convex curve enclosing all pixels of the lung. A convex set is a set in which two arbitrary points can be combined with a straight line that lies completely within the set. This convex casing 2 eliminates most of the irregularities 3, such as wedge-shaped perfusion defects, in the contour 1, which have arisen due to ventilation and perfusion defects in and near the lung edges. These irregularities can otherwise contribute to an incorrect adaptation of the lungs and thereby to incorrect diagnoses. Other image processing methods can also be used to eliminate the irregularities described above, such as “envelope” functions or the morphological operations “closure” or “dilatation” which are well known in the image processing technique.

Thereafter, the convex envelope of the ventilation image is adapted to the convex envelope of the perfusion image in step 202 to compensate for the patient not ingesting the exact same v ... ..--. -. -w-v. .- 10 15 20 25 30 35 516 851 f i :: = 9 position at the two registrations and the ventilation and perfusion images therefore do not show the lung from the same angle. This adaptation is preferably made according to Procrustes transformation, ie translation, rotation and magnification, after which an adapted image has been achieved.

This transformation is made as an image from the coordinates (x, y) to the coordinates (u, v) and is given by the function: u = cxcos (6) + cysin (9) + av = -cxsin (9) + cycos (6) + b which thus has four parameters. In order to be able to estimate the optimal parameters, a way is needed to assess how well the adaptation has been and this is done with the help of the covariance R, which is given by: E2 (V (X, y) -V. ,,) (P (X, y) -Pm) where V (x, y) is the intensity of the ventilation image at the point (x, y), V¿ is the average intensity of the ventilation image, P (x, y) is the intensity of the perfusion image at the point (x, y) , and Pm is the mean intensity of the perfusion image. The closer R is to the value 1, the better adjustment has been made, see Fig. 4. To facilitate the search for the optimal parameters, eg V5 and PB are used, which are functions of the contours, and which indicate the point sets that contain the points within the contour of each image, ie \ g (x, y) = 1 if (x, y) e Vs contour.

This method has the advantage that the adjustment does not depend on the intensity levels within the contours, which is why, for example, perfusion defects will not affect the adjustment. 10 15 20 25 30 35 516 85.1 10 The adjustment of the contours can also be done by using a temporary image material within the contours.

This image material is generated with a function that depends on the contour. For example, this function has the value 1 inside the contour and the value 0 outside. Other functions, such as the square of the distance between the contours, can also be used.

After the contours have been adapted to each other, the image material, which was within the contours before the adjustment, is transformed with the same transformation parameters used for the adjustment of the contours. In this way, the images have been adapted to each other in a robust and fast way.

Instead of Procrustes transformations, according to this invention it is possible to use other transforms, such as affine transformations, modifications of the Procrustes transformation where one can stretch or enlarge the contour differently at different angles, or non-parametric methods eg thin plate splines, be used. The parameters can also be determined in other alternative ways.

Then, in step 203, a difference image is formed from the matched images. This difference image, D, is the pixel-wise difference between the custom ventilation image and the perfusion image. This differential image is very well suited for use in disease assessment of a lung. This disease assessment can be performed by means of a neural network. In another embodiment of this invention, see Fig. 6, the intensities of the ventilation and perfusion images are scaled in step 603 before the difference image is generated in step 604. This is preferably done by scaling after the 50-95% intensity quantiles, but can also be done from the outside. the average intensity of the images. This intensity equalization compensates for defects caused by poor inhalation or for short registration times. n »anus 10 15 20 25 30 35. In another embodiment of this invention, see Fig. 7, pixels are removed in step 703, which pixels have a high intensity due to radioactive aerosol being stuck in the same place. These “hot-spots” result in large differences between the ventilation image and the perfusion image, which may be misinterpreted as emboli.

The intensity of the pixels in an image, which have an intensity higher than a 90-99% intensity quantile, is reduced or replaced with a lower value. The intensity of these points is interpolated with, for example, bicubic interpolation or Laplace interpolation.

In a preferred embodiment according to this invention, see Fig. 8, the contours are generated by thresholding.

The threshold is based on the average intensity in the images. It will be appreciated that the threshold value can be set either to a multiple, in the range of 0.5 to 3, of the mean, or according to an intensity quantile in the range of 10 to 90%.

Thereafter, the contour of the ventilation image is modified in step 802 so that the contour no longer includes any convex protrusions or contributions from the patient's neck and / or stomach, see Fig. 5 in which the upper lung scintigraphy image 13 shows a ventilation image of a lung 10 where a contribution 12 from the patient's neck is enclosed by the contour 11. The lower lung scintigraphy image 14 shows the same lung, but here the contour 11 has been changed so that the contribution 12 is no longer enclosed. These contributions arise by the radioactive aerosol also accumulating in the throat and / or stomach in the ventilation image. These neck contributions drastically change the shape of the contour and thereby also the convex casing formed therefrom, thereby making it difficult to properly adjust the lungs, as these contributions are often missing in the perfusion image. These contributions are connected to the lung via narrower segments.

These contributions can thus be removed if segments narrower than a predefined value are excluded. This idea is realized through the morphological operation erosion. This method corresponds to the removal of all internal parts 10 15 20 25 30 35 ~ - a Q nu 516 85.1 12 whereupon a ball with the diameter r cannot be placed so that it is completely inside the contours.

Alternatively, this step can be realized by "inflating" an elastic balloon inside the contour and replacing the contour with the largest balloon that fits inside the contour.

It should be understood that the contours can be generated by using some other edge detection, such as Laplace, Canny, Sobel and others.

According to another embodiment of this invention, the difference image is used as a basis for a disease assessment of the lung. In this assessment, it is practically impossible to analyze all pixels, since usually several sets of images are taken from different angles. In these cases, it is advantageous to instead make an analysis of various features, also called “features”, a so-called feature analysis. As a feature, the integral of the areas in the difference image that are underperfunded is preferably selected, ie the areas that have an intensity below a certain threshold value, a (which is, for example, equal to 0), in the difference image, ie Ai = L3 where Ai is the wound feature of the difference image in. If this integral exceeds a certain threshold value, c for one of the image pairs, the lung is judged to be diseased. In one embodiment of this invention, this assessment is performed by a neural network. The neural network is built in a multi-layered perceptron architecture with three layers; an entrance, a hiding and an exit warehouse. Networks of this kind and the like and their training are described in "Pattern recognition and Neural networks", Ripley BD, Cambridge University Press, ISBN O-521-46086-7.

It should be obvious to a person skilled in the art that other features can also be used and that the neural network can be structured differently. o o oo oooo oo o o .. o .oo. o oouo oo. oooo 10 15 oo coon oo I oo III 0 I 9 V oo oo o soo_ _ ooo 516 3.51 oonooo uoooo «oo .no uoooo oo oo: .. ooo 0 13 In another embodiment of this invention, the invention is realized by a system having: an X-ray or y-camera 20 (gamma camera) and an image processing means 21, for example a computer, which performs the method as described above, see Fig. 9.

In another embodiment according to this invention, the invention is realized by a memory medium, which has a computer program stored thereon, which performs several of the steps described above.

In another embodiment according to this invention, the invention is realized by an image processing device which performs one or more of the steps described above.

It will be appreciated that a variety of modifications to the above-described embodiments of the invention are possible within the scope of the invention as defined by the appended claims.

Claims (20)

10 15 20 25 30 35 516 s 51 f ”14 PATENT REQUIREMENTS A method of processing a digital ventilation image showing the airflow in a lung and a digital perfusion image showing the blood flow in the lung, characterized in that a set of adapted images is generated by adapting the images (202, 602, 702). , 804) to each other based on contours (1, 11) in the images, which contours (1, 11) are based on the intensity of the images. A method according to claim 1, wherein the adaptation (202, 602, 702, 804) is made by adapting the contour (1, 11) in at least one of the images to the contour (1, 11) in the second image and that the image material in said at least one image is adapted to the adapted contour (1, 11). The method of claim 2, wherein the matching (202, 602, 702, 804) comprises at least one of the steps of translating, rotating, enlarging and radially stretching the contour (1, 11) of an image. A method according to any one of the preceding claims, wherein the adaptation (202, 602, 702, 804) also includes a temporary image material within the contours (1, 11), which temporary image material is generated for an image by using a function of the contour of the image . A method according to any preceding claim, wherein the contours (1, 11) of the images are generated by threshold (801), which means that the intensity of the pixels of the images is compared with a threshold value. A method according to claim 5, wherein the threshold value is set to multiple, in the range 0.5 to 3, of the average value of the intensity in the image. A method according to any one of claims 1 to 4, wherein the contours (1, 11) are generated using an edge detection algorithm. u annu 10 15 20 25 30 35 5 1 6 s 5 1 - - Éjf 15 A method according to any preceding claim, wherein the contour (1, 11) of an image is replaced (201, 601, 701, 802) by its convex housing (2) before the adjustment is performed. A method according to any one of the preceding claims, further comprising the step of smoothing (802) the contour (1, 11) in at least one of the ventilation and perfusion images before the adjustment (202, 602, 702, 804) is performed, by exclude those parts of the contour (1, ll) whose mouth towards the enclosed lung volume is less than a preselected limit value, thereby removing any contribution (12) from the patient's other organs, such as its neck, from the image. A method according to any one of the preceding claims, wherein an intensity scaling (603, 704, 806) is performed on an image by calculating an intensity quantile in the range of 50-95% and then rescaling the intensities in the image based on this quantile should be the average intensity after the rescaling to compensate for defects, which are the result of, for example, poor inhalation or for short registration times. 11. ll. A method according to any one of the preceding claims, wherein the intensity of the pixels in an image which have an intensity higher than a 90-99% intensity quantile is reduced or replaced (703, 805) with a lower value. A method according to any one of the preceding claims, wherein a difference image is generated from the set of adapted images by subtracting (203, 604, 705, 807) one image from the other in the set of adapted images, which difference image is intended to form the basis for a disease assessment. A memory medium, which can be read by a computer (21) and on which is stored a computer program intended to process a digital ventilation image showing the air flow in a lung, and a digital perfusion image showing the blood flow in the lung, k characterized by the fact that the program is arranged to generate a set of custom images by adapting (202, 602, 702, 804). the images to each other based on contours (1, 11) in the images, which contours (1, 11) are based on the intensity of the images. A memory medium according to claim 13, wherein the program is arranged to generate the contours (1, 11) by threshold (801), which means that the intensity of the pixels of the images is compared with a threshold value. The memory medium of claim 13, wherein the program is arranged to generate the contours using an edge detection algorithm. A memory medium according to any one of claims 12 to 15, wherein the program is arranged to generate (203, 604, 705, 807) a difference image from the set of matched images by pixel-subtracting one image from the other in the set of matched images, which difference image is intended to form the basis for a disease assessment. An image processing device (21) for treating a lung image, the device having means for inputting a digital ventilation image showing the air flow in a lung, and a digital perfusion image showing the blood flow in a lung, characterized in that the device The arrangement is arranged to produce a set of adapted images by adapting the images (202, 602, 702, 804) to each other based on the contours of the images (1, 11), which are based on the intensity of the images. The image processing device (21) according to claim 17, wherein the device is arranged to generate the contours of the images (1, 11) by so-called threshold (801), i.e. to compare the intensity of the pixels of the images with a threshold value. , An image processing device (21) according to claim 17, which is arranged to generate the contours (1, 11) by means of an edge detection algorithm. An image processing device (21) according to any one of claims 17 to 19, which is arranged to generate a 0 .: vv- uv 516 851 17 difference image (203, 604, 705, 807) from the adapted ventilation and perfusion images by subtracting pixels one image from the other, which difference image is intended to form the basis for a disease assessment.