COMPUTER-IMPLEMENTED METHOD, SYSTEM AND COMPUTER PROGRAM PRODUCT FOR DETERMINING A VASCULAR FUNCTION OF A PERFUSION IMAGING SEQUENCE

Abstract

A computer-implemented method for determining a vascular function of a perfusion imaging sequence, includes the steps of: (i) receiving a perfusion imaging sequence comprising a voxel time series for a plurality of voxels; (ii) applying a trained classifier on the perfusion imaging sequence for receiving voxel-wise weights; (iii) receiving voxel-wise weights from the classifier; and (iv) determining the vascular function as the weighted sum of the voxel time series; wherein the classifier is trained by optimizing over the similarity between a predicted vascular function and a ground truth vascular function using a set of examples.

Claims

1.-15. (canceled)

16. A computer-implemented method for determining a vascular function of a perfusion imaging sequence, comprising the steps of: i. receiving a perfusion imaging sequence comprising voxel time series for a plurality of voxels; ii. applying a trained classifier on the perfusion imaging sequence for receiving voxel-wise weights; iii. receiving voxel-wise weights from the classifier; and iv. determining the vascular function as the weighted sum of the voxel time series; wherein the classifier is trained by optimizing over the similarity between a predicted vascular function and a ground truth vascular function using a set of examples.

17. The method according to claim 16, wherein the vascular function is the arterial input function or venous output function.

18. The method according to claim 16, wherein the voxel-wise weights are normalized.

19. The method according to claim 16, wherein the classifier is a first artificial neural network.

20. The method according to claim 19, wherein the first artificial neural network uses a spatial normalization, wherein the spatial normalization is a Softmax operation.

21. The method according to claim 19, wherein the training of the classifier further is an indirect training comprising creating a second artificial neural network and optimizing the second artificial neural network over the similarity between a predicted vascular function and a ground truth vascular function using the set of examples, wherein the second artificial neural network comprises the first artificial neural network.

22. The method according to claim 19, where the similarity is determined as the correlation between the predicted vascular function and the ground truth vascular function.

23. A system for determining a vascular function of a perfusion imaging sequence, the system comprising: means for receiving a perfusion imaging sequence, wherein the perfusion imaging sequence comprises voxel time series for a plurality of voxels; a trained classifier, wherein the classifier is configured to determine voxel-wise weights from the perfusion imaging sequence, wherein the classifier is trained by optimizing over the similarity between a predicted vascular function and a ground truth vascular function using a set of examples; means for applying the trained classifier on the perfusion imaging sequence for receiving voxel-wise weights and means for determining the vascular function as the weighted sum of the voxel time series.

24. The system according to claim 23, wherein the system further comprises: at least one processing unit; and means for performing a computer-implemented method for determining a vascular function of a perfusion imaging sequence, comprising the steps of: i. receiving a perfusion imaging sequence comprising voxel time series for a plurality of voxels; ii. applying a trained classifier on the perfusion imaging sequence for receiving voxel-wise weights; iii. receiving voxel-wise weights from the classifier; and iv. determining the vascular function as the weighted sum of the voxel time series; wherein the classifier is trained by optimizing over the similarity between a predicted vascular function and a ground truth vascular function using a set of examples.

25. The system according to claim 23, wherein the classifier is implemented as a first artificial neural network comprising a plurality of layers.

26. The system according to claim 25, wherein the plurality of layers comprises a last layer (Pvol) configured to output the voxel-wise weights, wherein the last layer is a spatial Softmax layer for normalizing the voxel-wise weights spatially.

27. The system according to claim 25, wherein the system further comprises a second artificial neural network, wherein the first artificial neural network is a part of the second artificial neural network and the classifier is trained indirectly by optimizing the second artificial neural network over the similarity between a predicted vascular function and a ground truth vascular function using the set of examples.

28. The system according to claim 25, wherein the first artificial neural network is implemented as a convolutional neural network, wherein the convolutional neural network comprises K convolutional layers L1 to LK, wherein each layer Lk comprising 23+k filters having a 3?3?3 kernel.

29. The system according to claim 28, wherein the convolutional neural network further comprises an extra convolution layer (Lout) behind the K convolutional layers with only one filter.

30. A computer program product for determining a vascular function of a perfusion imaging sequence comprising computer readable instructions for causing the system according to claim 23 to execute a computer-implemented method for determining a vascular function of a perfusion imaging sequence, comprising the steps of: i. receiving a perfusion imaging sequence comprising voxel time series for a plurality of voxels; ii. applying a trained classifier on the perfusion imaging sequence for receiving voxel-wise weights; iii. receiving voxel-wise weights from the classifier; and iv. determining the vascular function as the weighted sum of the voxel time series; wherein the classifier is trained by optimizing over the similarity between a predicted vascular function and a ground truth vascular function using a set of examples.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention is described in detail with reference to the examples shown in the drawings, in which the following is shown:

[0055] FIG. 1 a schematic view of an embodiment;

[0056] FIG. 2 a schematic view of an embodiment with higher detail to the classifier.

DETAILED DESCRIPTION OF VARIOUS EMBODIMENTS

[0057] FIG. 1 depicts a schematic view of an embodiment of the suggested method. At first, a perfusion imaging sequence 1 is provided. The perfusion imaging sequence 1 comprises a sequence of perfusion images, each image comprising three spatial dimensions. In alternate embodiments, the perfusion images may comprise only two spatial dimensions.

[0058] Each image is divided in elements of homogenous size, voxels. For example, each image comprises 256?256?2 voxels. Although larger or smaller images in any direction may be possible. Preferably, the perfusion imaging sequence 1 comprises between 10 and 120 images.

[0059] Preferably, the perfusion imaging sequence 1 is taken in between 40 and 120 seconds. Thus, the perfusion imaging sequence 1 comprises between 0.08 and 3 images per second.

[0060] A trained classifier 2 is applied to the perfusion imaging sequence 1. The classifier is preferably trained by optimization through backpropagation of the similarity between a predicted vascular function, i.e. the output, to a ground truth vascular function, i.e. the real vascular function of a set of examples. The ground truth vascular function may be determined by a human rater or another reliable determination process.

[0061] Applying the perfusion imaging sequence 1 to the classifier 2 results in voxel-wise weights 3. In other words, the classifier 2 outputs a weight 3 for each voxel of each image of the perfusion imaging sequence 1.

[0062] All voxel time series are weighted and added up. Thus, the weighted sum 4 of the voxel time series is determined. The weighted sum 4 gives the vascular function 5, which is the function of the intensity of a contrast agent within the blood over time.

[0063] FIG. 2 shows a schematic view of the classifier 2. In this embodiment, the classifier 2 is a convolutional neural network. As described above, the classifier receives the perfusion imaging sequence 1. The perfusion imaging sequence 1 comprises a time series of images x.sub.1 to x.sub.T. From these images, the classifier 2 determines voxel-wise weights for adding the weighted voxels to the weighted sum in the time series representing the vascular function y.sub.pred.

[0064] The classifier comprises K convolutional layers 6 (L.sub.1 to L.sub.K). Each Layer L.sub.k comprises 2.sup.(3+k) filters, one for every input channel of the input images or the preceding layer. Thus, the first layer L.sub.1 comprises 2.sup.4=16 filters, the second layer L.sub.2 comprises 2.sup.5=32 filters and so on.

[0065] The classifier 2 further comprises an extra convolutional layer Lout with only one filter. The extra convolutional layer Lout maps the convolutional layers to a single probabilistic volume.

[0066] The last layer of the classifier 2 is a layer for normalizing the voxel-wise weights spatially P.sub.vol. The last layer may preferably be a Softmax layer implementing a Softmax operation for spatially normalizing the voxel-wise weights. The normalized weights are output to means for determining the weighted sum 4 and thus, the vascular function 5.