Image data processing apparatus and method, a model training apparatus and a training method

1. An image data processing apparatus comprising:

processing circuitry configured to:

receive first medical image data;

receive second medical image data and/or stored image feature data;

extract a first image feature data set from the first medical image data;

extract a second image feature data set from the second medical image data and/or from the stored image feature data;

obtain style data;

extract a part of the first image feature data set and a part of the second image feature data set;

synthesize the part of the first image feature data set and the part of the second image feature data set to obtain a synthesized image data set; and

input the synthesized image data set and the style data to an artificial intelligence (AI) network to generate synthesized image data, wherein

the synthesized image data comprises a hybrid image that includes at least one anatomical feature from the first feature image data and at least one anatomical feature from the second feature image data, and

the synthesized image data set is obtained by performing an arithmetic operation using two or more of content factors Ca, Cb, and Cc, the two or more of content factors Ca, Cb, and Cc being included in the first image feature data set and/or the second image feature data set.

2. The image data processing apparatus according to claim 1 , wherein the processing circuitry is configured to perform a disentanglement process with respect to the first medical image data to extract the first image feature data set and/or to perform a disentanglement process with respect to the second medical image data to extract the second image feature data set.

3. The image data processing apparatus according to claim 2 , wherein the disentanglement process also comprises generating the style data.

4. The image data processing apparatus according to claim 1 , wherein at least one of a) or b):

a) at least one content factor of the two or more of content factors Ca, Cb, and Cc is representative of at least one anatomical feature or pathology; or

b) the first image feature data set and/or the second image feature data set comprises output of a trained model.

5. The image data processing apparatus according to claim 1 , wherein the style data is determined based on at least one of the first medical image data or the second medical image data and/or is sampled from a Gaussian or other distribution.

6. The image data processing apparatus according to claim 5 , wherein the synthesized image data is generated using a trained model that is trained by minimising adversarial loss and/or by incentivising consistency of background image and/or by incentivising correct classification of anatomical features of interest, pathologies of interest or other attributes, and/or by incentivising consistency of background content prior and after a noise injection.

7. The image data processing apparatus according to claim 1 , wherein the processing circuitry is configured to generate a mixed feature data set based on at least part of the first image feature data set and at least part of the second image feature data set.

8. The image data processing apparatus according to claim 7 , wherein the processing circuitry is further configured to:

generate the mixed feature data set by replacing at least one feature included in the first image feature data set with a corresponding feature from the second image feature data set.

9. The image data processing apparatus according to claim 7 , wherein the processing circuitry is further configured to:

generate the synthesized image data using a trained model, wherein the trained model is trained to generate the synthesized image data based on the mixed feature data set and the style data.

10. The image data processing apparatus according to claim 9 , wherein the trained model comprises a neural network that is configured to receive the mixed feature data set and the style data as inputs and to provide the synthesized image data as an output.

11. The image data processing apparatus according to claim 7 , wherein the processing circuitry is further configured to add noise data to the mixed feature data set.

12. The image data processing apparatus according to claim 11 , wherein the processing circuity is configured to generate the mixed feature data by replacing at least one feature included in the first image feature data set with a corresponding replacement feature from the second image feature data set, and the adding of noise data comprises adding the noise at a location of the replacement feature.

13. The image data processing apparatus of claim 1 , wherein the synthesized image data comprises spatial or other transformation of at least some of the first medical image data and/or at least some of the second medical image data, and/or at least some of the first image feature data set and/or the at least some of the second image feature data set.

14. The image data processing apparatus according to claim 1 , wherein the synthesized image data is generated using a trained model that is trained using a discriminator and a classifier, the discriminator is configured to discriminate between synthesized image data and image data and/or feature data used in generating the synthesized image data, and the classifier is used to identify anatomical features of interest, pathologies of interest or other attributes.

15. The image data processing apparatus according to claim 1 , wherein the style data is associated with a selected imaging modality and the generating of the synthesized image data is such that the synthesized image data represents an image that has an appearance consistent with being obtained using the selected imaging modality.

16. The image data processing apparatus according to claim 1 , wherein the style data is representative of at least one aspect of appearance of a medical image.

17. The image data processing apparatus according to claim 1 , wherein at least one of:

a) the processing circuitry is configured to generate the synthesized image data as a training data set for training a machine learning model;

b) the processing circuitry is configured to generate a plurality of synthesized image data sets and the generating of the synthesized image data sets comprises adding at least one pathology to image data derived from normal, non-pathological data sets thereby to provide additional pathological examples in a set of training data sets;

c) the synthesized image data set represents a selected patient or other subject and represents a modified version of the first medical image data or the second medical image data, and at least one of the first medical image data or second medical image data are obtained by scanning the patient or other subject.

18. An image data processing method, comprising:

receiving first medical image data;

receiving second medical image data and/or stored image feature data;

extracting a first image feature data set from the first medical image data,

extracting a second image feature data set from the second medical image data and/or from the stored image feature data;

obtaining style data;

extracting a part of the first image feature data set and a part of the second image feature data set;

synthesizing the part of the first image feature data set and the part of the second image feature data set to obtain a synthesized image data set; and

input the synthesized image data set and the style data to an artificial intelligence (AI) network to generate synthesized image data, wherein

the synthesized image data comprises a hybrid image that includes at least one anatomical feature from the first image feature data and at least one anatomical feature from the second feature image data, and

the synthesized image data set is obtained by performing an arithmetic operation using two or more of content factors Ca, Cb, and Cc, the two or more of content factors Ca, Cb, and Cc being included in the first image feature data set and/or the second image feature data set.

19. A model training apparatus comprising:

processing circuitry configured to:

receive a first image feature data set and a second image feature data set;

generate a mixed feature data set from the first image feature data set and the second image feature data set; and

train a model to generate a synthesized image data set from the mixed feature data set and style data, wherein

the synthesized image data comprises a hybrid image that includes at least one anatomical feature from the first feature image data and at least one anatomical feature from the second feature image data,

the first image feature data is extracted from a first medical image data,

the second image feature data is extracted from a second medical image data, and

the synthesized image data set is generated by performing an arithmetic operation using two or more of content factors Ca, Cb. and Cc, the two or more of content factors Ca, Cb, and Cc being included in the first image feature data set and/or the second image feature data set.

20. A training method, comprising:

receiving a first image feature data set and a second image feature data set;

generating a mixed feature data set from the first image feature data set and the second image feature data set; and

training a model to generate a synthesized image data set from the mixed feature data set and style data, wherein

the synthesized image data comprises a hybrid image that includes at least one anatomical feature from the first feature image data and at least one anatomical feature from the second feature image data,

the first image feature data is extracted from a first medical image data,

the second image feature data is extracted from a second medical image data, and

the synthesized image data set is generated by performing an arithmetic operation using two or more of content factors Ca, Cb, and Cc, the two or more of content factors Ca, Cb, and Cc being included in the first image feature data set and/or the second image feature data set.

21. An image data processing apparatus comprising:

processing circuitry configured to:

receive first medical image data;

receive second medical image data and/or stored image feature data;

obtain a first image feature data set from the first medical image data;

obtain a second image feature data set from the second medical image data and/or from the stored image feature data;

obtain style data; and

generate synthesized image data based on the first image feature data set, the second image feature data set, and the style data, wherein

the synthesized image data is generated using a trained model that is trained by minimising adversarial loss and/or by incentivising consistency of background image and/or by incentivising correct classification of anatomical features of interest, pathologies of interest or other attributes, and/or by incentivising consistency of background content prior and after a noise injection.