WO2025074145A1 - A computer-implemented method of identification of ischemia areas in the left ventricle - Google Patents

Abstract

A computer-implemented method (100) of identification regions of ischemia areas in the left ventricle, the method comprising: receiving (105) a CT Coronary Angiogram scan data representing a 3D volume of a region of anatomy that includes a pericardium; segmenting (110) the CT scan data to obtain a cardiac anatomical model; applying (110) a cardiac anatomical model to a machine-learning algorithm to identify a region of ischemia, while the machine-learning model is trained on a data sets comprising a spatially aligned CT and perfusion data.

Description

A COMPUTER-IMPLEMENTED METHOD OF IDENTIFICATION OF ISCHEMIA AREAS IN THE LEFT VENTRICLE

The present invention relates to a computer-implemented method of identification of ischemia areas in the left ventricle.

Ischemic heart disease (I HD), also known as coronary artery disease (CAD) or coronary heart disease (CHD), refers to a condition where there is a reduced blood supply to the heart muscle, typically due to the narrowing or blockage of the coronary arteries. The primary cause of the narrowing in the coronary arteries is the buildup of atherosclerotic plaques, which are deposits comprising cholesterol, lipid-rich substances, cellular waste products, calcium, and fibrin. Over time, these plaques can harden or rupture, leading to reduced or blocked blood flow, respectively.

Patients referred for ischemia diagnostics should be examined without undue delay, especially in emergency situations. One of the leading current techniques is FFR or CT-FFR, where results are interpreted by experienced radiologists who are not permanently available. Another technique evaluates blood flow to the heart muscle (myocardium) by doing tests CT Perfusion of the heart - OTP, Heart SPECT), Positron emission tomography (PET) of the heart muscle or Cardiac perfusion in magnetic resonance (OMR).

These are costly and time-consuming techniques; prolonged time to make the decision about implementation of an adequate medical procedure - consequently, it may also result in the patient's death in the absence of a prompt confirmation of life-threatening conditions.

Therefore in the field of identification and size assessment of ischemia areas there is a need for a fast and non-invasive method of identification of regions within the left ventricle with a low perfusion. However, obtaining perfusion data is laborious and time consuming, while the alternative invasive techniques, including FFR, pose additional risks to patients.

EP3277169A1 is disclosing systems and methods for using patient specific anatomical models and physiological parameters to estimate perfusion of a target tissue to guide diagnosis or treatment of cardiovascular disease. One method includes receiving a patient-specific vessel model and a patent-specific tissue model of a patient anatomy; extracting one or more patient-specific physiological parameters (e.g. blood flow, anatomical characteristics, image characteristics, etc.) from the vessel or tissue models for one or more physiological states of the patient; estimating a characteristic of the perfusion of the patient-specific tissue model (e.g., via a trained machine learning algorithm) using the patient-specific physiological parameters; and outputting the estimated perfusion characteristic to a display.

US11576639B2 discloses a method, system, device and medium for determining a blood flow velocity in cardiac arteries. An example method includes receiving a 3D model of the vessel, which is reconstructed based on X-ray angiography images of the vessel. The method further includes specifying a segment of the 3D model by a start landmark and a termination landmark. Moreover, the method includes determining the blood flow velocity based on length of the segment and perfusion time for the segment by normalizing the blood flow velocity to correspond to a cardiac cycle. The method presents high accuracy in calculating blood flow velocity and requires no additional modalities other than the original X-ray angiogram sequences used to visualize coronary arteries.

WO2019220417A1 discloses a method and system for automatically generating and analyzing fully quantitative pixel-wise myocardial blood flow and myocardial perfusion reserve maps to detect ischemic heart disease using cardiac perfusion magnetic resonance imaging. It is a computer- implemented method for automatically generating a fully quantitative myocardial blood flow map, comprising: receiving myocardial perfusion magnetic resonance imaging (MRI) images and arterial input function (AIF) MRI images; correcting a motion of a heart in the myocardial perfusion MRI images and the AIF MRI images, thereby obtaining motion corrected myocardial perfusion MRI images and motion corrected AIF images; correcting an intensity of the motion corrected myocardial perfusion MRI images and an intensity of the motion corrected AIF images, thereby obtaining surface coil intensity corrected MRI images and surface coil intensity corrected AIF images; using the surface coil intensity corrected MRI images and the surface coil intensity corrected AIF images, determining time-signal intensity characteristics and segmenting a left ventricle myocardial tissue region; and generating the myocardial blood flow map using the motion corrected myocardial perfusion MRI images, the left ventricle myocardial tissue region segmentation and the time-signal intensity characteristics.

Embodiments disclosed in US10080613B2 include a system for determining cardiovascular information for a patient. The system may include at least one computer system configured to receive patient-specific data regarding a geometry of the patient's heart, and create a three- dimensional model representing at least a portion of the patient's heart based on the patient-specific data. Then at least one computer system may be further configured to create a physics-based model relating to a blood flow characteristic of the patient's heart and determine a fractional flow reserve within the patient's heart based on the three-dimensional model and the physics-based model.

EP3600045A1 discloses a myocardial ct perfusion image synthesis, that relates to image processing devices and related methods. The image processing device comprises a data input for receiving spectral computed tomography volumetric image data organized in voxels. The image data comprises a contrast-enhanced volumetric image of a cardiac region in a subject's body and a baseline volumetric image of that cardiac region, e.g. a virtual non-contrast image, wherein the contrast-enhanced volumetric image conveys anatomical information regarding coronary artery anatomy of the subject. The device comprises a flow simulator for generating, or receiving as input, a three-dimensional coronary tree model based on the volumetric image data and for simulating a coronary flow based on the three-dimensional coronary tree model. The device comprises a perfusion synthesis unit for generating a perfusion image representative of a blood distribution in tissue at least one instant in time taking at least the baseline volumetric image and said coronary flow simulation into account.

The present invention is a computer-implemented method according to claim 1. The method relates to identification of ischemic regions in the left ventricle, the method comprising: receiving a CT Coronary Angiogram scan data representing a 3D volume of a region of anatomy that includes a pericardium; segmenting the CT scan data to obtain a cardiac anatomical model; applying a cardiac anatomical model to a machine-learning algorithm to identify a region of ischemia, while the machine-learning model is trained on a data sets comprising a spatially aligned CT and perfusion data test results being CTP, CMR, PET or SPECT as input.

Preferably the method according to the invention is characterized by a cardiac anatomical model that is comprising: left ventricle, papillary muscles, cardiac arteries, aorta, ostia points.

The present invention is also a computer-implemented method according to claim 3. The method relates to training a machine-learning algorithm to identify ischemia areas in the left ventricle, comprising the following steps: obtaining cardiac 3D CT images and a 3D myocardial perfusion spatial data; extracting a set of features characterizing the left ventricle in a form of a 3D CT mask and 3D perfusion mask with characteristic anatomical points identified, including detectable subset of: valves, papillary muscles, apex cordis, ostia points, chordae tendineae, trabeculae carneae and the left ventricle outer line shape used as landmarks. The next step involves applying a continuous transformation to a 3D perfusion mask to arrive at cardiac CT data augmented by data extracted from a 3D perfusion mask to create a spatially aligned CT and perfusion data; and training a machine-learning model on data sets comprising a spatially aligned CT and perfusion data.

Preferably, in the method according to the invention a machine learning model includes one or more of the following architectures: 3D U-Nets, V-Nets, or ResNets, CNN including DRL or GAN tailored for 3D image segmentation.

Preferably in the method according to the invention a set of features characterizing to the left ventricle 3D mask includes extracting at least one of the following features: cardiac arteries, lumen geometry, or stenosis.

Preferably in the method according to the invention a continuous transformation is a landmark based 3D morphing image deformation.

Preferably in the method according to the invention the 3D myocardial perfusion spatial data include at least one data category from the group comprising: CT Perfusion of the heart - CTP; Myocardial perfusion scintigraphy (Heart SPECT); Positron emission tomography (PET) of the heart muscle; Cardiac perfusion in magnetic resonance (CMR).

Preferably the method according to the invention further comprises denoising a 3D myocardial perfusion spatial data.

Preferably in the method according to the invention the machine-learning model trained is a GAN structure with a generator that accepts a 3D cardiac CT image and outputs a predicted perfusion map values while an input comprises a voxel-based representation of cardiac CT 3D images, with two steps. The first step using a multichannel 2D GAN nets processing UV 2D representation of the data, generating a base output ischemia structure, with U-Net type architecture as generator and PatchGAN type structure as discriminator, and second step using 3D volumetric GAN to refine the base output converted back to 3D representation in areas along the centerlines of arteries. This second phase structure has the following layers: Initial Layer which is a convolutional layer with filters adjusted for 3D data, Initial Graph Convolutional Layers processing the input (latent vector + left ventricle + arteries) and Graph Deconvolutional/Up-sampling Layers, and a discriminator distinguishes between real and generated ischemia regions.

Preferably in the method according to the invention the myocardial perfusion spatial data are CT Perfusion of the heart - CTP, the method comprising the following steps: obtaining CTP data based on TDC (Time-Density Curve) data from each voxel; preparing two 3D masks per case with the following data: Myocardial blood flow (MBF), directly reflecting the amount of blood reaching the myocardial tissues; Mean Transit Time (MTT), reflecting the time contrast medium takes to traverse a specific tissue region, indicative of compromised blood flow, such as in ischemic areas. Further the method includes a step of noise reduction using Gaussian kernel 5x5x5 followed by data segmenting by grouping segments of points with average values below threshold TMBF and TMTT for MBF and MTT masks respectively, calculated as relative measures where TMBF = 0.75 * MaxMBF and TMTT = 0.75 * MaxMTT, MaxMBF and MaxMTT being calculated after applying Gauss smoothing. Next, it involves applying a mask transformation to transpose the 3D perfusion masks onto 3D cardiac-CT mask with: detecting a left ventricle in both data sets; detecting characteristic points in both data sets, comprising at least three characteristic points chosen from a group comprising: valves, papillary muscles, apex cordis, ostia points, chordae tendineae, trabeculae carneae and the left ventricle outer line shape; applying a continuous transformation to perfusion mask, where the input cardio-CT is augmented with MBF and MTT values from the perfusion mask.

Advantages of the Invention: Reduces the need for additional invasive FFR or perfusion imaging by predicting ischemia from standard cardiac CT 3D images. Speeds up diagnosis and potentially leads to earlier interventions. Provides an opportunity for real-time or near-real-time predictions in clinical settings. The innovative concept of using artificial intelligence for detecting ischemia areas based on CT Coronary Angiogram image analysis will enable obtaining results with a very high degree of accuracy (specificity and sensitivity in real time). The present invention will significantly speed up the diagnostic and decision-making process and reduce the costs associated with the potential elimination of the need to implement invasive diagnostic procedures and will provide immediate help in diagnosis and guidance on the optimal treatment strategy.

The present invention is a modern Al application designed as SaMD - Software as Medical Device based on the known relationship between the degree of ischemia and stenosis of the coronary vessels observed in angiography. It is a unique tool based on multidimensional neural networks analyzing Cardiac-CT imagery, including artery geometry, connections and stenosis, in result indicating ischemic areas. It provides an efficient qualitative and quantitative assessment of cardiac Ischemia on the basis of CT Coronary Angiogram examination without the need for further diagnostics precise evaluation of the size of the myocardial ischemic area. It provides a realistic 3D model of the examined heart and coronary vessels allowing the detectors to perform a thorough analysis of the current state and deep understanding of the anatomy and pathophysiology of the given patient.

The present invention is described in a preferred embodiments in relations to the drawings, where:

Fig. 1 shows a block diagram of the method according to the invention; Fig. 2 shows a block diagram of a training method;

Fig. 3a and 3b show a structure of training data set, used in base processing and fine tuning of the output result;

Fig. 4 shows a graphical representation of the ischemic area detected by the method according to the invention.

A block diagram 100 of a method according to claim 1 of the present invention is shown in Figure 1. The method comprises a series of steps performed by the processor using data input into a memory via computer system interfaces.

The computer system interfaces comprise a standard data communication interface including a computer network interface, a data storage medium interface, etc.

In the first step 105, the computer system processor receives CT scan data representing a 3D volume of an anatomical region including a pericardium. This data is a typical dataset received in a non-invasive computed tomography method in which a number of X-ray scans are combined into a dataset representing features of the scanned object, preferably an anatomical region including a pericardium.

In the next step 110, the processor performs a segmentation step comprising segmenting the CT scan data to obtain an anatomical model of the heart. This process includes segmenting the left ventricle with papillary muscles, aorta, ostia points, left ventricle artery centerlines with lumen and the shape of the interior of the arteries.

In the next step 120, the processor applies the anatomical model of the heart to a machine learning algorithm to identify the ischemic region while training the machine learning model on a dataset comprising spatially aligned CT and perfusion data input.

The result of such application is a 3D model of the area affected by ischemia with two sub-regions marked ischemic core and penumbra.

In the preferred embodiment of the invention a cardiac anatomical model comprises: left ventricle, cardiac arteries, aorta, ostia points, papillary muscles.

The machine learning model used in step 120 is a result of the training method of claim 3. A block diagram 200 of a method according to claim 3 is shown in Figure 2.

As seen in Figure 2 a computer-implemented method for training a machine learning algorithm for identifying areas of ischemia in the left ventricle, comprises the number of steps described in detail below. The first step 205 is a step of obtaining 3D cardiac CT images and 3D myocardial perfusion spatial data.

In the next step 210 a set of features characterizing the left ventricle is extracted in the form of a 3D CT mask and a 3D perfusion data is obtained from perfusion test dataset, which for the whole training dataset can be uniformly one of the following: CMR (Cardiac Magnetic Resonance), CTP (CT Perfusion), PET (Positron Emission Tomography) or SPECT (Single Photon Emission Computed Tomography), followed by a step 215 of transforming a 3D myocardial perfusion mask into a 3D cardiac CT mask by detecting left ventricular bitmasks by characteristic anatomical points.

In the following step 220 the processor applies a continuous transformation to a 3D perfusion mask to obtain cardiac CT data which is augmented by data extracted from a 3D perfusion mask to produce spatially aligned CT and perfusion data.

A continuous transformation to a 3D perfusion mask is done using landmark based 3D morphing image deformation which produces cardiac CT data which is augmented by data extracted from a 3D perfusion mask, maintaining a relative spatial alignment is crucial to produce effects of the invention.

Data structure can be seen in Figure 3a and 3b. The sample data comprises left ventricle 3D structure, the centerlines and shape-wise lumen data, along with stenosis markings. The data is prepared for two neural network groups - first, preprocessing, for convoluted 2D nets operating on multiple channels and the second, post-processing, for the volumetric 3D generative nets. The data is generated by the first group and refined by the second.

Data structure used for the pre-processing phase involves Multi-channel Stacking: The 2D representations of the ventricle surface, centerlines, lumen, shape, and stenosis markings are stacked as separate channels to form a multi-channel 2D input. Data structure used for the postprocessing phase using volumetric 3D networks.

The final step 225 comprises training a machine learning model on data sets comprising spatially aligned CT and perfusion data.

In preferred embodiments of machine learning models include one or more of the following architectures: 3D U-Nets, V-Nets, or ResNets, CNN including DRL or GAN tailored for 3D image segmentation.

In particular preferred embodiment features characterizing to the to the left ventricle 3D mask includes extracting at least one of the following features: cardiac arteries, lumen geometry, or stenosis.

A 3D myocardial perfusion spatial data include at least one data category from the group comprising: CT Perfusion of the heart - CTP; Myocardial perfusion scintigraphy (Heart SPECT); Positron emission tomography (PET) of the heart muscle; Cardiac perfusion in magnetic resonance (CMR).

In a preferred embodiment the training method further comprises denoising a 3D myocardial perfusion spatial data. This step involves assigning specific to given perfusion test perfusion value vector to each voxel, then using using 5x5x5 kernel gaussian filter to perfusion values detecting the highest perfusion values in each vector position, normalizing the whole perfusion mask so that the areas of the highest oxidation value have value 1 and the least 0 (in each perfusion vector position), clustering the points being inside minimal connected convex sets including points below thresholds Ti , calculated as relative measures for each vector position i, where Ti = 0.75 * Maxi. In one of preferred embodiments the machine-learning model trained is a GAN structure that comprises a generator that accepts a 3D cardiac CT image and outputs a predicted ischemia area while an input comprises a voxel-based representation of cardiac CT 3D images.

The GAN data flow is as follows: GAN generator taking cardiac CT 3D images as input along with anatomy data (arteries centerlines and shape, stenosis) as specified in Fig. 3a, and GAN discriminator taking both anatomical data representation and ischemia areas - either real or synthesized.

The process is done in two phases - base ischemia area generation on a 2D UV mapping representation and refinement done on 3D volumetric representation done along the arteries' centerlines.

Phase 1 : for generating the base ischemia 3D marking, taking 2D representation of the input as specified in Fig. 3a GAN structure:

GENERATOR:

U-Net Architecture:

® The U-Net architecture is designed to shuttle low-level information directly across the network, which is especially useful for image-to-image translation problems where the input and output share a lot of low-level information.

® The U-Net has skip connections between each layer i in the encoder and layer n-i in the decoder, where n is the total number of layers. These skip connections concatenate activations from layer i to layer n-i.

Encoder-Decoder Structure:

® Encoder: The encoder consists of a series of layers that progressively downsample the input.

® Decoder: The decoder upsamples the encoded information to produce the output.

DISCRIMINATOR:

PatchGAN structure:

® The discriminator is designed to penalize structure at the scale of image patches ® The PatchGAN discriminator used to classify if each NxN patch in an image is real or synthesized. This discriminator operates convolutionally across the image, and the responses are averaged to provide the final output of the discriminator.

® This design effectively models the image as a Markov random field, assuming independence between pixels separated by more than a patch diameter.

Then the output image is converted back to 3D representation by using the opposite of the transformation 3D->2D UV described in Fig. 3a. Phase 2: The next step uses volumetric 3D GAN nets to post-process the output along the 3D lines indicated by artery lines,, taking 3D representation of the input as specified in Fig. 3b GAN GAN Structure for 3D Mesh Representation:

Generator (G):

® I nput: Latent Vector: A random noise vector or latent representation that provides variability for the GAN, Left Ventricle Structure: The 3D mesh or representation of the left ventricle., Arteries with Lumen Structure: The encoded structure of the arteries and their lumen

® Graph Convolutional Layers: Processing the input to generate a coarse version of the 3D mesh, performing convolutions on the graph/mesh structure, aggregating information from neighboring vertices to update the attributes of each vertex. Allowing the network to capture local geometric structures, including spatial relationships between the left ventricle, arteries, and lumen

® Graph Deconvolutional/Up-sampling Layers: Refining and up-sampling the mesh to its final resolution.

® Output: A 3D mesh representing the left ventricle with potential ischaemia areas.

Discriminator (D):

® Input: A 3D ischemia mesh (either real or generated by G).

® Graph Convolutional Layers: Processing the mesh to capture its geometric and topological features.

® Pooling/Down-sampling Layers: Down-sampling the mesh to capture higher-level features.

® Fully Connected Layers: Processing the down-sampled mesh to make a decision.

® Output: A scalar value indicating whether the input ischemia mesh is real or generated.

In this embodiment a discriminator distinguishes between real and generated ischemia area maps.

In yet another preferred embodiment wherein the myocardial perfusion spatial data are CT Perfusion of the heart - CTP, the method comprising the following steps: obtaining CTP data based on TDC (Time-Density Curve) data from each voxel; preparing two 3D masks per case with the following data:

Myocardial blood flow (MBF), directly reflecting the amount of blood reaching the myocardial tissues.

Mean Transit Time (MTT), reflecting the time contrast medium takes to traverse a specific tissue region, indicative of compromised blood flow, such as in ischemic areas. The next step is the step of noise reduction as CTP data. Promising effects are achieved by using using 5x5x5 kernel gaussian filter to MBV and MTT values respectively, normalizing the whole perfusion mask so that the areas of the highest oxidation value have value 1 and the least 0 (in MBV and MTT layers individually), clustering the points being inside to minimal connected convex sets including points below thresholds TMBF and TMTT for MBF and MTT masks respectively, calculated as relative measures where TMBF = 0.75 * MaxMBF and TMTT = 0.75 * MaxMTT, MaxMBF and MaxMTT being calculated after applying Gauss smoothing, and marking such clusters as ischemia areas (in MBV and MTT layers individually).

The following step if the one of applying a mask transformation to transpose the 3D perfusion masks onto 3D CT Coronary Angiogram mask with: detecting a left ventricle in both data sets; detecting characteristic points in both data sets, comprising at least one characteristic points chosen from a group comprising: valves, papillary muscles and chordae tendineae, trabeculae carneae.

The applying a continuous transformation to perfusion mask follows, where the input cardio-CT is augmented with MBF and MTT values from perfusion mask.

Figure 4 shows a graphical representation of ischemia area detected by the method according to the invention. The area 401 is delimited from the remaining structure with a dark black line to improve clarity of drawing only, invention can use any type of presentation means, like grayscaling, coloring, applying specific pattern etc, to indicate the ischemia area detected.

Claims

Patent claims

1 . A computer-implemented method (100) of identification regions of ischemia areas in the left ventricle, the method comprising: a) receiving (105) a CT scan data representing a 3D volume of a region of anatomy that includes a pericardium; b) segmenting (110) the CT scan data to obtain a cardiac anatomical model; c) applying (115) a cardiac anatomical model to a machine-learning algorithm to identify a region of a low perfusion, while the machine-learning model is trained on a data sets comprising a spatially aligned CT and perfusion data.

2. The method according to claim 1 wherein a cardiac anatomical model is comprising: left ventricle, cardiac arteries, aorta, ostia points, papillary muscles.

3. A computer-implemented method (200) of training a machine-learning algorithm to identify ischemia areas in the left ventricle, comprising the following steps: obtaining (205) cardiac 3D CT images and a 3D myocardial perfusion spatial data; extracting (210) a set of features characterizing to the left ventricle in a form of a 3D CT mask and 3D perfusion mask; followed by transposing (215) a 3D myocardial perfusion mask onto a cardiac 3D CT mask by detection left ventricle bitmasks by characteristic anatomical points; followed by applying (220) a continuous transformation to a 3D perfusion mask to arrive at cardiac CT data augmented by data extracted from a 3D perfusion mask to create a spatially aligned CT and perfusion data; training (225) a machine-learning model on data sets comprising a spatially aligned CT and perfusion data.

4. The method according to claim 3 wherein a machine learning model includes one or more of the following architectures: 3D U-Nets, V-Nets, or ResNets, CNN including DRL or GAN tailored for 3D image segmentation.

5. The method according to any of claims 3 or 4 wherein a set of features characterizing to the to the left ventricle 3D mask includes extracting at least one of the following features: cardiac arteries, lumen geometry, or stenosis.

6. The method according to claim 3 or 5 wherein a continuous transformation is a landmark based 3D morphing image deformation.

7. The method according to any of claims from 3 to 6 wherein 3D myocardial perfusion spatial data include at least one data category from the group comprising: CT Perfusion of the heart - CTP; Myocardial perfusion scintigraphy (Heart SPECT); Positron emission tomography (PET) of the heart muscle; Cardiac perfusion in magnetic resonance (CMR).

8. The method according to claim 7 wherein it further comprises denoising a 3D myocardial perfusion spatial data.

9. The method according to any of claims from 2 to 8 wherein the machine-learning model trained is a GAN structure that comprises a generator that accepts a 3D cardiac CT image and outputs a predicted ischemia area while an input comprises a voxel-based representation of cardiac CT 3D images.

10. The method according to claim 9 wherein the GAN data flow is as follows: a) GAN generator taking cardiac CT 3D images as input along with anatomy data and GAN discriminator taking both anatomical data representation and ischemia areas - either real or synthesized; b) phase 1 : for generating the base ischemia 3D marking, taking 2D UV representation of the input GAN structure:

GENERATOR with

U-Net Architecture:

® The U-Net architecture is designed to shuttle low-level information directly across the network, which is especially useful for image-to- image translation problems where the input and output share a lot of low-level information;

® The U-Net has skip connections between each layer I in the encoder and layer n-i in the decoder, where n is the total number of layers;

Encoder-Decoder Structure:

® Encoder: The encoder consists of a series of layers that progressively downsample the input

® Decoder: The decoder upsamples the encoded information to produce the output;

DISCRIMINATOR with

PatchGAN structure:

• The discriminator is designed to penalize structure at the scale of image patches;

• The PatchGAN discriminator used to classify if each NxN patch in an image is real or synthesized; then the output image is converted back to 3D representation by using the opposite of the transformation 3d->2D UV; c) phase 2: with volumetric 3D GAN nets to post-process the output along the 3D lines indicated by artery lines

GAN Structure for 3D Mesh Representation:

Generator (G):

® I nput: Latent Vector: A random noise vector or latent representation that provides variability for the GAN, Left Ventricle; Structure: The 3D mesh or representation of the left ventricle; Arteries with Lumen Structure: The encoded structure of the arteries and their lumen;

® Graph Convolutional Layers: Processing the input to generate a coarse version of the 3D mesh, performing convolutions on the graph/mesh structure, aggregating information from neighboring vertices to update the attributes of each vertex, allowing the network to capture local geometric structures, including spatial relationships between the left ventricle, arteries, and lumen;

® Graph Deconvolutional/Up-sampling Layers: Refining and up-sampling the mesh to its final resolution;

® Output: A 3D mesh representing the left ventricle with potential ischaemia areas;

Discriminator (D):

® Input: A 3D ischemia mesh (either real or generated by G).

® Graph Convolutional Layers: Processing the mesh to capture its geometric and topological features;

® Pooling/Down-sampling Layers: Down-sampling the mesh to capture higher-level features;

® Fully Connected Layers: Processing the down-sampled mesh to make a decision;

® Output: A scalar value indicating whether the input ischemia mesh is real or generated.

11. The method according to any of claims 3 to 10 wherein the myocardial perfusion spatial data are CT Perfusion of the heart - CTP, the method comprising the following steps: obtaining CTP data based on TDC (Time-Density Curve) data from each voxel; preparing two 3D masks per case with the following data:

Myocardial blood flow (MBF), directly reflecting the amount of blood reaching the myocardial tissues;

Mean Transit Time (MTT), reflecting the time contrast medium takes to traverse a specific tissue region, indicative of compromised blood flow, such as in ischemic areas; a step of noise reduction; data segmenting by grouping segments of points with average values below threshold TMBF and TMTT for MBF and MTT masks respectively, calculated as relative measures; obtaining 3D continuous value maps by smoothing procedure; applying a mask transformation to transpose the 3D perfusion masks onto 3D cardiac-CT mask with: detecting a left ventricle in both data sets; detecting characteristic points in both data sets, comprising at least one characteristic points chosen from a group comprising: valves, papillary muscles and chordae tendineae, trabeculae carneae; applying a continuous transformation to perfusion mask, where the input cardio- CT is augmented with MBF and MTT values from perfusion mask.