DE102010028382A1 - Method for processing tomographic image data from X-ray computed tomography investigation of liver for recognition of liver tumor, involves performing iterative classification, and calculating image mask from last probability image - Google Patents

Abstract

The method involves loading (S1) an image data set including multitude of image pixels, and performing normalization (S2) of the image data set. A feature vector is calculated for each of the image pixels with image features based on the properties of the image pixels, and an iterative classification is performed (S3) by adding feature vectors of each image pixel with probability characteristics. The iterative classification is stopped based on stopping criterion, and an image mask is calculated (S4) from a last probability image, where each image pixel is assigned unique to a certain class. An independent claim is also included for a computer system for processing tomographic image data from X-ray CT investigation of a liver.

Description

The invention relates to a method and a computer system for processing tomographic image data from an X-ray CT examination of an examination subject by calculating feature vectors for the image pixels containing properties of the image pixels and / or their surroundings, classifying the image pixels on the basis of the calculated feature vectors Probability considerations and determination of a candidate image mask that uniquely associates each image pixel with a particular class.

Similar methods are generally known in the context of image processing. By way of example, the German Offenlegungsschrift DE 10 2006 017 291 A1 directed. This document discloses a method and a computer system for segmenting a tomographic volume data set from image values of an X-ray CT examination of an examination subject comprising a learning phase and a recognition phase, wherein in the learning phase a Probabilistic Boost Tree (= PBT) based on prior knowledge by application is trained on different material areas and comparison of the results of the PBT with the previously determined material classes and determined in the recognition phase property vectors of the material areas and assigned by applying the trained PBT on the material areas each material area a particular class of material.

Such methods can be used on the one hand in the field of material testing and baggage control. By appropriate parameterization of programs in a computer system, this computer system can be used in conjunction with image data from an examination of a patient with a CT system. In particular, such systems are used for the diagnosis and follow-up of liver tumors, as well as for the planning of appropriate interventions. Since liver tumors usually differ in their density hardly from the surrounding liver tissue, the patient is usually intravenously administered an X-ray contrast medium prior to recording. Subsequently, several CT volumes are recorded, showing different phases of contrast enhancement in the liver. Tumors and other lesions show a specific and different from the parenchyma enrichment behavior, so that they can be easily detected and classified in these recordings. For diagnosis, intervention planning and follow-up a precise segmentation of the lesions in question, ie a point-by-point subdivision of the image into lesions and background, is essential. Since the liver in the shots described can extend over several hundred layers, manual segmentation is not practicable due to time constraints, therefore, a computer system is required for this, which performs this procedure automatically.

Since in pictures of the venous phase, ie when the contrast medium is mainly in the liver venous system, most of the lesions present hypodense, previous methods are often based on adaptive threshold methods. Other approaches apply clustering techniques such as "k-means" to previously filtered images to separate lesions and background. However, both approaches require a good contrast between the lesion and the surrounding tissue, which is not always the case even in contrast medium exposures. The actual contrast depends heavily on the timing of the recording and the individual perfusion of the liver.

The approach closest to the method of the invention uses the AdaBoost machine learning method to classify points in the liver as belonging to a lesion or background. Here, two classifiers are trained, one classifier for detecting large lesions and the other for detecting small lesions. Both are used to process a new image and their results are merged. In this regard, the publication "Ensemble segmentation using AdaBoost with application to liver lesion extraction from a CT volume" by SHIMIZU A., NARIHIRA T., FURUKAWA D., KOBATAKE H., NAWANO S., SHINOZAKI K, in The MIDAS Journal, 2008 , referenced. A disadvantage of this method is that here an arbitrary separation of the points is made, without this separation would be motivated from the data. It is also unclear how many such separations should be made.

Another weakness of previous methods is the lack of consideration of different contrasting of the data. In part, a histogram equalization is performed, but only visually provides improved contrast while at the same time destroying the intensity distribution shape. In part, images are also normalized by subtracting their mean from all intensities and dividing them by their standard deviation. However, since the intensity of the parenchyma is significantly more influenced by the contrast agent than that of the lesion, in particular in hypodense lesions, normalization may shift the gray scale range of the lesion further than desired.

It is therefore an object of the invention to find an improved method for processing tomographic image data from an X-ray CT examination of an examination subject and recognition of predetermined classified objects. Furthermore it is intended to propose a computer system implementing this method, in particular for use in the diagnostic field.

This object is solved by the features of the independent claims. Advantageous developments of the invention are the subject of the subordinate claims.

The inventors have recognized that in an automatic, probability-based classification of tomographic image data from an X-ray CT examination of an examination object, it is particularly favorable if, in addition to the probability-based classification, the image pixel takes an iterative classification due to its feature properties and output of a probability image mask the feature vectors with properties on the basis of the respectively last determined probability image mask are used for the classification and thus iteratively calculates new probability image masks and the last calculated probability image mask is again used to calculate a candidate image mask which assigns each image pixel to a particular class.

• – Laden eines ersten Bilddatensatzes, bestehend aus einer Vielzahl von Bildpixel mit ersten Bildwerten,
• – Normierung der ersten Bildwerte auf einen vorgegebenen Standard,
• – Berechnung eines Merkmalsvektors für jedes Bildpixel mit Bildmerkmalen auf der Basis der Eigenschaften der Bildpixel und/oder deren Umgebung,
• – erste Klassifikation der Bildpixel in Form eines Wahrscheinlichkeitsbildes auf der Basis der Merkmalsvektoren mit Berechnung der Wahrscheinlichkeit der Zugehörigkeit des jeweiligen Bildpixels zu einer Klasse von Bildpixel aus mindestens zwei unterschiedlichen Klassen durch ein zuvor trainiertes und auf Wahrscheinlichkeitswerten basierendes Klassifikationsverfahrens mit den gleichen Merkmalsvektoren und Erstellung eines Wahrscheinlichkeitsbildes aus den bildpixelbezogenen Wahrscheinlichkeitswerten,
• – iterative Klassifikation durch:
• – Ergänzung der Merkmalsvektoren jedes Bildpixels mit Wahrscheinlichkeitsmerkmalen auf der Basis der Wahrscheinlichkeitswerte der Bildpixel und/oder deren Umgebung mit dem zuletzt ermittelten Wahrscheinlichkeitsbild,
• – Durchführung mindestens einer weiteren Klassifikation durch Bewertung der ergänzten Merkmalsvektoren und
• – Ausgabe eines neuen Wahrscheinlichkeitsbildes mit Wahrscheinlichkeitswerten für die Zugehörigkeit der Bildpixel zu mindestens einer Klasse,
• – Abbruch der iterativen Klassifikation nach einem vorgegebenen Abbruchkriterium und Ausgabe des letzten Wahrscheinlichkeitsbildes als Ergebnis, und
• – Berechnung einer Kandidatenbildmaske aus dem letzten Wahrscheinlichkeitsbild, wobei jedes Bildpixel eindeutig einer bestimmten Klasse zugeordnet wird.

Accordingly, the inventors propose a method for processing tomographic image data from an X-ray CT examination of an examination subject, comprising the method steps:

• Loading a first image data set consisting of a plurality of image pixels with first image values,
• - normalization of the first image values to a predetermined standard,
• Calculation of a feature vector for each image pixel with image features on the basis of the properties of the image pixels and / or their surroundings,
• - First classification of the image pixels in the form of a probability image on the basis of the feature vectors with calculation of the probability of belonging the respective image pixel to a class of image pixels from at least two different classes by a previously trained and based on probability values classification method with the same feature vectors and creating a probability image from the image pixel-related probability values,
• - iterative classification by:
• Supplementing the feature vectors of each image pixel with probability features on the basis of the probability values of the image pixels and / or their surroundings with the last-determined probability image,
• Carrying out at least one further classification by evaluating the supplemented feature vectors and
• Outputting a new probability image with probability values for the association of the image pixels with at least one class,
• Abort the iterative classification according to a predetermined abort criterion and output the last probability image as a result, and
• Calculation of a candidate image mask from the last probability image, wherein each image pixel is uniquely assigned to a particular class.

To avoid unnecessary arithmetic operations, it may be advantageous if, prior to the normalization of the image values, a pre-segmentation of an image area is carried out on the basis of a predetermined property and the further method takes place exclusively on this pre-segmented image area.

It is particularly favorable if the normalization of the image values takes place by transformation of the image values in such a way that the histogram of the transformed image values subsequently corresponds to a predetermined standard histogram, that is to say standardization is performed by histogram adaptation.

It is also advantageous if at least one filter is applied to the determined probability image mask to even out the probability image mask after the iterative classification. In this case, one-, two- or three-dimensional filters as well as linear or non-linear filters can be used. Accordingly, such filters can alternatively or additionally also be used for equalization on the candidate mask.

• – Grauwert des betrachteten Bildpixels,
• – minimaler Grauwert der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – maximaler Grauwert der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – mittlerer Grauwert der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Median der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Kontrast der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Spannweite der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Varianz der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Schiefe der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung,
• – Gradienten der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung in verschiedene Richtungen,
• – Laplace-Operator der Grauwerte der Bildpixel einer vorgegebenen zweidimensionalen Umgebung,
• – Differenz aus dem Mittelwert einer zwei- oder dreidimensionalen Umgebung des betrachteten Bildpixels und dem Mittelwert aller Punkte des vorgegebenen Bereiches, und
• – 3D-Haar-Merkmale der Grauwerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung. Der Begriff Grauwert steht hier selbstverständlich im Sinne eine gemessenen Intensitätswertes beziehungsweise eines sonstigen Bildwertes der tomographischen Bilddaten.

With regard to the properties of the feature vectors used in the first approach, the inventors propose that these should describe at least one property of the list below:

• Gray value of the considered image pixel,
• Minimum gray value of the image pixels of a given two- or three-dimensional environment,
• Maximum gray value of the image pixels of a given two- or three-dimensional environment,
• Mean gray value of the image pixels of a given two- or three-dimensional environment,
• Median of the gray values of the image pixels of a given two- or three-dimensional environment,
• Contrast of the gray values of the image pixels of a given two- or three-dimensional environment,
• Span of the gray values of the image pixels of a given two- or three-dimensional environment,
• Variance of the gray values of the image pixels of a given two- or three-dimensional environment,
• Skewness of the gray values of the image pixels of a given two- or three-dimensional environment,
• Gradients of the gray values of the image pixels of a given two- or three-dimensional environment in different directions,
• Laplace operator of the gray values of the image pixels of a given two-dimensional environment,
• Difference between the mean value of a two-dimensional or three-dimensional environment of the observed image pixel and the average of all points of the given region, and
• 3D hair features of the gray levels of the image pixels of a given two- or three-dimensional environment. Of course, the term gray value here means a measured intensity value or another image value of the tomographic image data.

• – Wahrscheinlichkeitswert des betrachteten Bildpixels,
• – mittlerer Wahrscheinlichkeitswert der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung einschließlich dem betrachteten Bildpixel,
• – mittlerer Wahrscheinlichkeitswert der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung ausschließlich dem betrachteten Bildpixel,
• – Median der Wahrscheinlichkeitswerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung, und
• – Gauss-gewichtete Summe der Wahrscheinlichkeitswerte der Bildpixel einer vorgegebenen zwei- oder dreidimensionalen Umgebung.

Furthermore, it is proposed that the supplemented feature vectors additionally or exclusively describe at least one property of the following list determined from the probability image:

• Probability value of the considered image pixel,
• Mean probability value of the image pixels of a given two- or three-dimensional environment including the considered image pixel,
• Mean probability value of the image pixels of a given two- or three-dimensional environment excluding the observed image pixel,
• Median of the probability values of the image pixels of a given two- or three-dimensional environment, and
• Gauss-weighted sum of the probability values of the image pixels of a given two- or three-dimensional environment.

As a classification method can be used in the context of the invention, for example, a PBT process (PBT = Probabilistic Boosting Tree), an AdaBoost method, or a support vector machine method.

In addition, to calculate the candidate image mask, a threshold or value range for the probability of each image pixel in the probability image mask may be set, followed by class assignment. In this case, it may be particularly favorable if the user of the method according to the invention is offered the option of individually setting this limit value or this value range, preferably with simultaneous observation of the resulting candidate image mask or its superposition with the first image data record.

Furthermore, the first image data set in conjunction with the candidate image mask, in particular in the form of a color overlay, can be output on a screen.

As the first image data set, a two-dimensional sectional image, a three-dimensional sequence of such sectional images or an original three-dimensional image data set can be used.

As already mentioned, the method is particularly suitable for material examination, in particular in the context of quality control or baggage inspection.

In addition to the method described above, the invention also proposes a computer system with a memory for recording computer programs, in particular in conjunction with a CT system, which has computer programs in the memory of the computer system which execute the method according to the invention during operation.

In this case, a patient can also be provided as the examination object, whereby, if the examination is limited to a predetermined area, this area can be an organ, in particular the liver. In particular, we proposed here that the classes are healthy organ tissue on the one hand and tumor on the other hand.

In the following the invention with reference to a preferred embodiment with reference to the figures will be described in more detail, with only the necessary features for understanding the invention features are shown. The following reference numerals are used: B: CT image; C1: CT system; C2: first X-ray tube; C3: first detector; C4: optional X-ray tube; C5: optional detector; C6: gantry housing; C7: patient; C8: examination couch; C9: system axis; C10: computer system; C11: contrast agent applicator; C12: ECG lead; PBT1-PBT3: Probalistic Boosting Tree; Prg ₁ prg _n : computer programs; S1: preprocessing; S1.1: segmentation; S1.2: standardization; S2: image pixelwise classification; S2.1: generation of image features; S2.2: first classification; S3: iteration; S3.1: generation of the probability features; S3.2: iterative classification; S4: post-processing; S4.1: filtering; S4.2: candidate generation; S5: output; W1-W3: Probability Pictures.

They show in detail:

1 : a schematic representation of the sequence of the method according to the invention,

2 a representation of a PBT cascade including the calculated probability images and

3 A CT system with computer system for carrying out the method according to the invention for processing tomographic CT image data.

The following description of the invention relates, without limitation to the general public, to the recognition of a liver tumor in a liver of a patient. Here, the normal liver tissue for any first class of material and the tumor tissue for any second class of material.

It is particularly advantageous if the detection and the segmentation of the liver lesions are carried out simultaneously by using a classifier to decide for each point in the liver whether it belongs to a lesion or to the parenchyma. The entry into the classification system forms a CT image of the liver with venous contrast enhancement. In detail, this is divided into the 1 schematically illustrated method in the steps preprocessing S1 with segmentation S1.1 - here the liver and standardization S1.2 of the image values, image pixel-wise classification S2 with the calculation of the image features S2.1 and formation of image pixel-related feature vectors and the first classification S2.2, by which a probability image for the presence of a particular class is calculated. Subsequently, the iteration S3 is carried out with the method steps of possibly calculating the probability features S3.1 and supplementing the feature vectors or replacing the previous corresponding vector values with these probability features and the classification S3.2 taking into account the last determined probability image. Finally, the post-processing S4, having at least one filtering S4.1 and a subsequent candidate generation S4.2, takes place, for example, by using a probability threshold value. The last step S5 is the output of the result, for example in the form of the candidate image itself or superposition or marking of the CT image with the candidate image or a candidate mask resulting therefrom.

The individual steps of such a process are as follows:

- Segmentation of the liver

The liver is segmented by a method known per se. For example, an automatic segmentation method can be used as described in the document Haibin Ling, S. Kevin Zhou, Yefeng Zheng, Bogdan Georgescu, Michael Suehling, Dorin Comaniciu, IEEE CVPR, pp. "Hierarchical, learning-based automatic liver segmentation". 1-8, 2008 , is described. All further steps can then only take place within the liver presentation. This restricts the number of points to be classified, thus speeding up the process. In addition, large areas of feature space are excluded from the classification from the outset, which significantly reduces the intra-class variance of the problem and thus simplifies the work of the classifier.

- Standardization of image values

The intensities of the points in the liver representation can now with the method according to the document "Nonrigid registration of joint histograms for intensity standardization in magnetic resonance imaging", Jäger, F., Hornegger, J., IEEE Trans Med Imaging, pp. 137-150, Issue 1, 2009 to which a reference data record is to be aligned. For this purpose, its histogram is calculated and registered elastically on the histogram of the reference data set. This results in a transformation that assigns a new value to each gray value in the image. This transformation is then applied to the image.

- Feature calculation

For each pixel in the liver representation a number of features are calculated. These form for each pixel a feature vector - sometimes referred to as an input pattern or example - which is then used for the classification. The spectrum of possible features is very broad. In this context, in addition to the intensity or gray value of the pixel itself, features from gray value statistics such as minimum and maximum gray value, mean, median, contrast, span, variance, skewness, respectively calculated in 2D and 3D environments of various sizes around the Point, continue gradient features in 2D and 3D, as well as 3D hair features of different scales.

- Iterative classification

For the classification step, for example, comes the Probabilistic Boosting Tree (PBT), as in the publication "Probabilistic boosting tree: Learning discriminative models for classification, recognition, and clustering", Zhuowen Tu, Proceedings of the Tenth IEEE International Conference on Computer Vision - Volume 2, pp. 1589-1596, 2005 , or in the publication US 2007/0053563 A1 used in combination with an AdaBoost method and a threshold learning method. This PBT, prior to its application to new data, must first be trained on a set of data sets using the machine learning procedure of the same name, in which the lesions to be found have been segmented manually or by means of semi-automatic algorithms.

If the finished trained PBT is applied to a new data set, it calculates the probability of the presence of a lesion, in the form of a probability image, for each feature vector and therefore for each point.

If each pixel is classified for itself, then adjacent pixels are implicitly assumed to be independent. However, for contiguous objects spanning multiple pixels, this assumption is incorrect, so the system described here uses not only a PBT but a cascade of iteratively executed PBTs. In this case, the first PBT is trained with the image features described above. Subsequently, the lesion probability for each pixel is calculated with it. The resulting probability image is then used to calculate further features that are appended to the previous feature vectors. Other features that can be used are, for example: the probability value of the pixel itself, the mean value, a 2D median, a 3D median, the mean value without the point itself and / or the Gauss weighted sum of the environment in 2D and 3D. With these extended feature vectors, the next PBT is trained on it and the entire process is carried out analogously in one or multiple iterations.

The application of this system to CT image data is the same and is in the 2 shown schematically. For each pixel of a CT image B feature vectors are calculated with image features based on the image information and passed to the first Probabilistic Boost Tree PBT1. From the output probability image additional probability features are calculated, which form the input to the second Probabilistic Boost Tree PBT2 together with the image features as supplemental feature vectors. This Probabilistic Boost Tree PBT2. generates a new probability image, with the properties of which the previous probability features in the feature vector are supplemented or replaced in order to be fed to the next Probabilistic Boost Tree, here the PBT3. In this way, the vectors go through the whole cascade with PBTs, at the end for each point a final probability image for the selected class, here a lesion, is output.

- Post processing

In principle, in order to generate candidate masks for lesions from the classification step from the classification step, it is sufficient to define a threshold value for the probability over which the pixels are regarded as a lesion. With this threshold, the sensitivity of the method against its specificity can be weighed. However, in order to get a smooth result, the image is previously processed with a median filter that removes discrete points with very different probability values and then performs a morphological opening operation to remove further small false positive detections.

Using a Probabilistic Boosting Tree to classify the pixels has several advantages. Due to its hierarchical structure, it works according to the divide & conquer principle. In this way he can detect much more complex problems and larger intra-class variances than other classifiers. The use of several classifiers to reduce the variance in the data is thus superfluous or takes place implicitly in the training of the tree. Compared to the pure AdaBoost method, it has the advantage that it can be much faster with the same performance, since the classification of a pixel under certain circumstances may not have to go through the entire tree.

By stacking several PBTs in a single cascade, the classification result becomes smoother and more robust. The lesion probability of a point calculated by a PBT thus influences the decision of the next PBT for the neighboring points. Strong fluctuations of the probabilities in a small space are thus avoided. Visually, this is in the 2 recognizable, in each of which the calculated probability images W1 to W3 are shown below each PBT.

The standardization of the image values, ie the measured intensities, compensates for differences in the intensity distribution between images by adapting the distribution to a reference distribution. As opposed to rigid methods, such. B. intensity reduction of the entire image by a fixed value, or a mere normalization with average and standard deviation different gray scale ranges are treated separately, it can be ensured that different tissue types are adjusted separately. For example, if the parenchyma is brighter in an image than in the reference dataset, the lesion you are looking for is not, a rigid procedure would still change the intensity of the lesion. In the method described here, ideally, the intensity of the lesion is treated detached and therefore not changed. This achieves a very good comparability of the data records.

The 3 finally shows an exemplary CT system C1 with which the inventive method can be performed. The CT system C1 has a first tube / detector system with an X-ray tube C2 and an opposite detector C3. Optionally, this CT system C1 may also have a second X-ray tube C4 with an opposite detector C5. Both tube / detector systems are located on a gantry, which is arranged in a gantry housing C6 and rotates during the scanning about a system axis C9. The patient C7 is located on a movable examination couch C8, which is pushed either continuously or sequentially along the system axis C9 through the scanning field located in the gantry housing C6, whereby the attenuation of the X-ray radiation emitted by the X-ray tubes is measured by the detectors.

During the measurement, the patient C7 can be injected with a contrast agent bolus with the aid of a contrast agent applicator C11, so that blood vessels can be better recognized or a perfusion measurement can be carried out. For cardio recordings, the heart activity can also be measured using an ECG C12 cable and an ECG gated scan can be performed.

The control of the CT system is carried out with the aid of a computer system C10, in which there are computer programs Prg ₁ to Prg _n , which can also carry out the method according to the invention described above. Additionally, about this. Computer system C10 also made the output of image data.

Alternatively, however, the computer system C10 may also be used separately, receiving and processing only detector data or image data of a CT examination, and processing them within the scope of the invention described above.

Overall, therefore, a very robust segmentation and classification of material classes, in particular of liver lesions despite varying contrasting and variation of the tumors in terms of shape, size and texture is achieved by the method according to the invention or a corresponding image processing system.

It is understood that the abovementioned features of the invention can be used not only in the respectively specified combination but also in other combinations or in isolation, without departing from the scope of the invention.

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 102006017291 A1 [0002]
• US 2007/0053563 A1 [0034]

Cited non-patent literature

• "Ensemble segmentation using AdaBoost with application to liver lesion extraction from a CT volume" by SHIMIZU A., NARIHIRA T., FURUKAWA D., KOBATAKE H., NAWANO S., SHINOZAKI K, in The MIDAS Journal, 2008 [0005]
• Haibin Ling, S. Kevin Zhou, Yefeng Zheng, Bogdan Georgescu, Michael Suehling, Dorin Comaniciu, IEEE CVPR, pp. "Hierarchical, learning-based automatic liver segmentation". 1-8, 2008 [0031]
• "Nonrigid registration of joint histograms for intensity standardization in magnetic resonance imaging", Jäger, F., Hornegger, J., IEEE Trans Med Imaging, pp. 137-150, Issue 1, 2009 [0032]
• Zhuowen Tu, Proceedings of the Tenth IEEE International Conference on Computer Vision - Volume 2, p. 1589-1596, 2005 [0034]

Claims (17)

Method for processing tomographic image data (B) from an X-ray CT examination of an examination object (C7), comprising the following method steps: 1.1. Loading a first image data set consisting of a multiplicity of image pixels with first image values, 1.2. Normalization of the first image values to a given standard, 1.3. Calculating a feature vector for each image pixel with image features based on the properties of the image pixels and / or their environment, 1.4. first classification of the image pixels in the form of a probability image on the basis of the feature vectors with calculation of the probability of belonging the respective image pixel to a class of image pixels from at least two different classes by a previously trained and based on probability values classification method with the same feature vectors and creating a probability image from the image pixel related probability values, 1.5. iterative classification by: 1.5.1. Supplementing the feature vectors of each image pixel with probability features on the basis of the probability values of the image pixels and / or their surroundings with the last determined probability image (W1, W2), 1.5.2. Carrying out at least one further classification by evaluating the supplemented feature vectors and 1.5.3. Output of a new probability image (W1, W2) with probability values for the association of the image pixels with at least one class, 1.6. Abort the iterative classification according to a predetermined abort criterion and output the last probability image (W3) as a result, and 1.7. Calculation of a candidate image mask from the last probability image (W3), wherein each image pixel is uniquely assigned to a specific class. Method according to the preceding Patent Claim 1, characterized in that, prior to the normalization of the image values, a pre-segmentation of an image area on the basis of a predetermined property is carried out and the further process is carried out exclusively on this pre-segmented image area. Method according to one of the preceding claims 1 to 2, characterized in that the normalization of the image values is performed by transforming the image values such that thereafter the histogram of the transformed image values corresponds to a predetermined standard histogram. Method according to one of the preceding claims 1 to 3, characterized in that for equalization of the probability image mask after the iterative classification at least one filter is applied to the determined probability image mask. Method according to one of the preceding claims 1 to 4, characterized in that for equalization of the candidate mask on this at least one filter is applied. Method according to one of the preceding claims 1 to 5, characterized in that the feature vectors describe at least one property of the list below: Gray value of the considered image pixel, minimum gray value of the image pixels of a given two- or three-dimensional environment, maximum gray value of the image pixels of a given two- or three-dimensional environment, average gray value of the image pixels of a given two- or three-dimensional environment, Median of the gray values of the image pixels of a given two- or three-dimensional environment, Contrast of the gray values of the image pixels of a given two- or three-dimensional environment, Range of the gray values of the image pixels of a given two- or three-dimensional environment, Variance of the gray values of the image pixels of a given two- or three-dimensional environment, Skewness of the gray values of the image pixels of a given two- or three-dimensional environment, Gradients of the gray values of the image pixels of a given two- or three-dimensional environment in different directions, Laplace operator of the gray values of the image pixels of a given two-dimensional environment, Difference between the mean value of a two- or three-dimensional environment of the observed image pixel and the mean value of all points of the given region, and 3D hair features the gray levels of the image pixels of a given two- or three-dimensional environment. Method according to one of the preceding claims 1 to 6, characterized in that the supplemented feature vectors describe at least one property of the following list determined from the probability image: probability value of the considered image pixel, mean probability value of the image pixels of a given two- or three-dimensional environment including the considered image pixel, mean probability value of the image pixels of a given two- or three-dimensional environment excluding the considered image pixel, median of the probability values of the image pixels of a given two- or three-dimensional environment, and Gauss weighted sum the probability values of the image pixels of a given two- or three-dimensional environment. Method according to one of the preceding claims 1 to 7, characterized in that as classification method one of the methods according to the following list is used: - PBT process (PBT = Probabilistic Boosting Tree), - AdaBoost procedure, - Support vector machine procedure. Method according to one of the preceding claims 1 to 8, characterized in that for the calculation of the candidate image mask, a limit or value range for the probability of each image pixel in the probability image mask is set, after which the class assignment takes place. Method according to one of the preceding claims 1 to 9, characterized in that the first image data set in conjunction with the candidate image mask, in particular in the form of a color overlay, is output on a screen. Method according to one of the preceding claims 1 to 10, characterized in that the first image data set is a two-dimensional image data set. Method according to one of the preceding claims 1 to 10, characterized in that the first image data set is a three-dimensional image data set. Method according to one of the preceding claims 1 to 12, characterized in that the method is carried out as a material examination, in particular in the context of a quality control or baggage inspection. Computer system (C10) having a memory for recording computer programs, in particular in conjunction with a CT system (C1), characterized in that in the memory of the computer system (C10) computer programs (Prg ₁ to Prg _n ) are stored, which in operation the Method according to one of the method claims 1 to 12 performs. Computer system according to the preceding patent claim 14, characterized in that a patient (C7) is provided as the examination subject. Computer system according to the preceding patent claim 15, characterized in that, when the examination is limited to a predetermined area, this area is an organ, in particular the liver. Computer system according to one of the preceding claims 14 to 16, characterized in that as classes at least also healthy organ tissue and tumor is used.

Images