DE102017203299B4 - System and method for cardioembolic stroke risk prediction based on medical images - Google Patents

Abstract

A method for stroke risk prediction based on medical images, comprising: extracting measurements of the left atrium (LA) and left atrial appendage (LAA) from medical image data of a patient; calculating derived metrics for the patient's LA and LAA using a patient-specific one Computational model of the cardiac function based on the measurements of the LA and LAA extracted (104) from the patient's medical image data; and calculating a stroke risk score for the patient based on the extracted LA and LAA measurements and the calculated (106) derived metrics for the patient's LA and LAA using a classifier based on the trained machine learning, the extracted (104 ) LA and LAA measurements and the calculated (106) derived metrics for the LA and LAA are entered as features into the classifier based on the trained machine learning.

Description

STATE OF THE ART

The present invention relates to patient-specific stroke risk prediction and, more particularly, to patient-specific cardioembolic stroke risk prediction based on medical images.

Stroke is the fifth leading cause of death and the leading cause of adult disability in the United States. There are 2 types of strokes, hemorrhagic and ischemic, with approximately 13% of strokes being hemorrhagic and 87% being ischemic. Ischemic strokes can in turn be embolic (-20%) or thrombotic (-80%) of the type. In an embolic stroke, a blood clot or fragment of plaque forms in the body (usually in the heart) and travels to the brain, where it blocks a small vessel. In a thrombotic stroke, the blood clot forms inside an artery that supplies blood to the brain.

Embolic strokes are more disabling than non-embolic strokes due to the occlusion of large intracranial arteries and larger ischemic brain volume. It has been estimated that 45 to 50% of embolic strokes occur in hearts with atrial fibrillation (Afib). Furthermore, it has been estimated that approximately 2.3-3.2 million people were affected with Afib in the United States in 2011, and based on epidemiological data, future projections of patients with Afib could exceed 12 million by 2050.

Current clinical practice regarding risk prediction for cardiac embolic stroke patients is extensive and includes detailed history, physical examination (including cardiac auscultation for heart murmurs and assessment for an irregular heartbeat), neuroimaging, electrocardiogram, and laboratory and echocardiographic data. However, the hazard indices derived from such assessments are limited to simpler statistical indicators with large variances, usually based on longitudinal clinical data. There are a number of disadvantages to this method of risk prediction, including the time plasticity (i.e. uncertainty) and the large range of these indices, as well as weak / reduced patient specificity.

From the US2011257545A1 an image-based symptomatic classification and cardiovascular risk assessment is known. the US2016058520A1 discloses a system and a method for patient-specific image-based simulation of artificial electrophysiology. the US2012022843A1 discloses a method and system for comprehensive patient-specific modeling of the heart. A clinical risk analysis for the occurrence of a stroke in people with atrial fibrillation, a common cardiac arrhythmia, emerges from the website "www.wikipedia.de" CHA2DS2-VASc Score "version dated September 19, 2016".

SUMMARY OF THE INVENTION

The present invention provides a method and system for patient-specific prediction of ischemic stroke risk. Embodiments of the present invention provide ischemic stroke risk prediction based on automated analysis of the left atrial (LA) appendage and left atrial appendage (LAA).

In one embodiment of the present invention, measurements of the left atrium (LA) and the left atrial appendage (LAA) are extracted from medical image data of a patient. Derived metrics for the patient's LA and LAA are calculated using a patient-specific computation model of cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data. A stroke risk rating for the patient is calculated based on the extracted LA and LAA measurements and the calculated derived metrics for the LA and LAA of the patient using a classifier based on the trained machine learning, with the extracted LA and LAA Measurements and the calculated derived metrics for the LA and LAA are entered as features of the classifier based on the trained machine learning.

In order to overcome the disadvantages, the invention proposes, according to a first aspect, a method for patient-specific stroke risk prediction on the basis of medical images, comprising:

Extracting measurements on the left atrium (LA) and left atrium appendage (LAA) from medical image data of a patient;

Computing derived metrics for the LA and LAA of the patient using a patient-specific computation model of the cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data; and

Calculating a stroke risk score for the patient based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a classifier based on trained machine learning, with the extracted LA and LAA Measurements and the calculated derived metrics for the LA and LAA are entered as features into the classifier, which is based on the trained machine learning.

• Segmentieren des LA in den medizinischen Bilddaten des Patienten.

According to a preferred development of the invention, a method can be provided wherein extracting the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient comprises:

• Segmenting the LA in the patient's medical image data.

• Erzeugen eines 3D-LA-Gitters aus medizinischen 3D-Bilddaten des Patienten.
• Erzeugen eines 3D-LA-Gitters aus medizinischen 3D-Bilddaten des Patienten kann umfassen:
  • Segmentierung mehrerer LA-Teile in den medizinischen Bilddaten unter Verwendung eines mehrteiligen Atriummodells; und
  • Erzeugen eines konsolidierten LA-Gitters aus den Segmenthirten mehreren Atriumteilen.

Segmenting the LA in the patient's medical image data can include:

• Generation of a 3D LA grid from medical 3D image data of the patient.
• Generating a 3D LA grid from 3D medical image data of the patient can include:
  • Segmenting multiple LA parts in the medical image data using a multi-part atrium model; and
  • Create a consolidated LA grid from the segment shepherds of several atrium parts.

• Erzeugen einer Folge von LA-Gittern aus medizinischen 4D (3D + Zeit)-Bilddaten des Patienten.

According to a preferred development of the invention, a method can be provided wherein segmenting the LA in the patient's medical image data comprises:

• Generate a sequence of LA grids from medical 4D (3D + time) image data of the patient.

• Extrahieren eines LA-Volumens, LAA-Volumens und einer Anzahl von LAA-Lappen auf der Basis des segmentierten LA, wobei das LA-Volumen, LAA-Volumen und eine Anzahl von LAA-Lappen als Merkmale in den auf maschinelles Lernen basierten Klassifikators eingegeben werden.

According to a preferred development of the invention, a method can be provided, wherein the extraction of the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient further comprises:

• Extracting an LA volume, LAA volume and a number of LAA lobes based on the segmented LA, with the LA volume, LAA volume and a number of LAA lobes being entered as features into the machine learning based classifier .

• Extrahieren von hämodynamischen Messungen für das LA und LAA aus den medizinischen Bilddaten des Patienten.

According to a further preferred development of the invention, a method can be provided, wherein the extraction of the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient further comprises:

• Extract hemodynamic measurements for the LA and LAA from the patient's medical image data.

According to a further preferred development of the invention, a method can be provided wherein the calculation of derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the heart function on the basis of the LA and LAA measurements obtained from the medical image data of the Patients extracted includes: simulating blood flow in the LA and LAA using the patient-specific computation model of cardiac function; and calculating the hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being input into the classifier based on trained machine learning.

• Berechnen von abgeleiteten hämodynamischen Parametern für mindestens mehrere Orte im LAA, das auf dem simulierten Blutstrom im LA und LAA basiert, wobei die abgeleiteten hämodynamischen Parameter ein oder mehrere Elemente aus der relativen Verweilzeit (RRT), Energieverlust, Druckverlustkoeffizient,
• Wandschubspannung (WSS) oder Oszillationsindex (OSI) enthalten.
• Das Berechnen von hämodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA kann umfassen:
  • Berechnen für mindestens mehrere Orte im LAA, eine statistische Kennzeichnung von mindestens einem Element aus Blutströmungsgeschwindigkeit oder Druck, auf der Basis des simulierten Blutstroms im LA und LAA.
  • Das Berechnen von abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines patientenspezifischen Berechnungsmodells der Herzfunktion auf der Basis der Messungen des LA und LAA, die aus den medizinischen Bilddaten des Patienten extrahiert wurden, kann ferner umfassen:
• Simulieren der elektrischen Signalausbreitung im LA unter Verwendung des patientenspezifischen Berechnungsmodell der Herzfunktion; und
• Berechnen von elektrophysiologischen Merkmalen für das LA und LAA aus der simulierten elektrischen Signalausbreitung im LA, wobei die berechneten elektrophysiologischen Merkmale als Merkmale in den Klassifikator eingegeben werden, der auf dem maschinellen Lernen basiert.

According to a further preferred development of the invention, a method can be provided wherein calculating the hemodynamic characteristics for the LA and LAA based on the simulated blood flow in the LA and LAA comprises:

• Calculation of derived hemodynamic parameters for at least several locations in the LAA, which is based on the simulated blood flow in the LA and LAA, the derived hemodynamic parameters being one or more elements from the relative residence time (RRT), energy loss, pressure loss coefficient,
• Wall shear stress (WSS) or oscillation index (OSI) included.
• Computing hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA can include:
  • Calculation for at least several locations in the LAA, a statistical identification of at least one element from blood flow velocity or pressure, on the basis of the simulated blood flow in the LA and LAA.
  • Computing derived metrics for the patient's LA and LAA using a patient-specific computation model of cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data may further include:
• Simulate electrical signal propagation in the LA using the patient-specific calculation model of the heart function; and
• Calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being entered as features into the classifier, which is based on machine learning.

According to a further preferred development of the invention, a method can be provided, wherein the extracted LA and LAA measurements, which are input as features into the classifier, which is based on machine learning, the LA volume, LAA volume and include the number of LAA lobes, and the calculated derived metrics that are entered as features into the classifier based on machine learning include one or more items from Relative Residence Time (RRT), Energy Loss, Pressure Loss Coefficient, Wall Shear Stress (WSS ), Oscillation Index (OSI), mean blood flow velocity or mean blood pressure, at one or more points in the LA and LAA.

• Berechnen der Schlaganfallrisikoeinstufung des Patienten auf der Basis der extrahierten LA- und LAA-Messungen und der berechneten abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines trainierten tiefen neuronalen Gitters.

According to another preferred development of the invention, a method can be provided wherein calculating a stroke risk score of the patient based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a classifier , which is based on machine learning, includes:

• Calculating the patient's stroke risk score based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a trained deep neural grid.

• Mittel zum Extrahieren von Messungen am linken Atrium (LA) und linken Atriumsanhang (LAA) aus medizinischen Bilddaten eines Patienten;
• Mittel zum Berechnen von abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines patientenspezifischen Berechnungsmodells der Herzfunktion auf der Basis der Messungen des LA und LAA, die aus den medizinischen Bilddaten des Patienten extrahiert wurden; und Mittel zum Berechnen einer Schlaganfallrisikoeinstufung für den Patienten auf der Basis der extrahierten LA- und LAA-Messungen und der berechneten abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines Klassifikators, der auf dem trainierten maschinellen Lernen basiert, wobei die extrahierten LA- und LAA-Messungen und die berechneten abgeleiteten Metriken für das LA und LAA als Merkmale in den Klassifikator eingegeben werden, der auf dem trainierten maschinellen Lernen basiert.

In order to overcome the disadvantages, the invention proposes, according to a second aspect, a device for patient-specific stroke risk prediction on the basis of medical images, comprising:

• Means for extracting measurements on the left atrium (LA) and left atrial appendage (LAA) from medical image data of a patient;
• Means for calculating derived metrics for the patient's LA and LAA using a patient-specific calculation model of cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data; and means for calculating a stroke risk score for the patient based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a classifier based on the trained machine learning, the extracted LA- and LAA measurements and the calculated derived metrics for the LA and LAA are entered as features into the classifier based on the trained machine learning.

• Mittel zum Segmentieren des LA in den medizinischen Bilddaten des Patienten.
• Messungen des linken Atriums (LA) und des linken Atriumsanhangs können als Merkmale in den Klassifikator eingegeben werden, der auf dem trainierten maschinellen Lernen basiert, umfassen ein LA-Volumen, LAA-Volumen und eine Zahl von LAA-Lappen, die auf der Basis des segmentierten LA bestimmt wird.
• Mittel zum Extrahieren von Messungen des linken Atriums (LA) und des linken Atriumsanhangs aus medizinischen Bilddaten eines Patienten können ferner umfassen:
  • Mittel zum Extrahieren von hämodynamischen Messungen für das LA und LAA aus den medizinischen Bilddaten des Patienten.

According to a further preferred development of the invention, a device can be provided, wherein the means for extracting the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient comprises:

• Means for segmenting the LA in the patient's medical image data.
• Measurements of the left atrium (LA) and left atrial appendage can be entered as features into the classifier, which is based on trained machine learning, include an LA volume, LAA volume, and a number of LAA lobes based on the segmented LA is determined.
• Means for extracting measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient can further comprise:
  • Means for extracting hemodynamic measurements for the LA and LAA from the patient's medical image data.

• Mittel zum Simulieren des Blutstroms im LA und LAA unter Verwendung des patientenspezifischen Berechnungsmodells der Herzfunktion; und
• Mittel zum Berechnen von hämodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA, wobei die hämodynamischen Merkmale in den Klassifikator eingegeben werden, der auf dem trainierten maschinellen Lernen basiert.
• Das Mittel zum Berechnen von hämodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA kann umfassen:
  • Mittel zum Berechnen von abgeleiteten hämodynamischen Parametern für mindestens mehrere Orte im LAA, das auf dem simulierten Blutstrom im LA und LAA basiert, wobei die abgeleiteten hämodynamischen Parameter ein oder mehrere Elemente aus der relativen Verweilzeit (RRT), Energieverlust, Druckverlustkoeffizient, Wandschubspannung (WSS) oder Oszillationsindex (OSI) enthalten.
  • Außerdem oder alternativ kann das Mittel zum Berechnen von hämodynamischen Merkmalen für das LA und LAA auf Basis des simulierten Blutstroms im LA und LAA umfassen:
    • Mittel zum Berechnen für mindestens mehrere Orte im LAA, eine statistische Kennzeichnung von mindestens einem Element aus Blutströmungsgeschwindigkeit oder Druck, auf der Basis des simulierten Blutstroms im LA und LAA.
    • Außerdem oder alternativ kann das Mittel zum Berechnen der abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines patientenspezifischen Berechnungsmodells der Herzfunktion auf der Basis der Messungen des LA und
    • LAA, die aus den medizinischen Bilddaten des Patienten extrahiert wurden, ferner umfassen:
      • Mittel zum Simulieren der elektrischen Signalausbreitung im LA unter Verwendung des patientenspezifischen Berechnungsmodells der Herzfunktion; und
      • Mittel zum Berechnen von elektrophysiologischen Merkmalen für das LA und LAA aus der simulierten elektrischen Signalausbreitung im LA, wobei die berechneten elektrophysiologischen Merkmale als Merkmale in den Klassifikator eingegeben werden, der auf dem maschinellen Lernen basiert.

According to a further preferred development of the invention, an apparatus can be provided, wherein the means for calculating derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the heart function on the basis of the LA and LAA measurements obtained from the medical Image data extracted from the patient includes:

• Means for simulating the blood flow in the LA and LAA using the patient-specific computation model of the heart function; and
• Means for calculating hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being input into the classifier which is based on the trained machine learning.
• The means for calculating hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA may include:
  • Means for calculating derived hemodynamic parameters for at least several locations in the LAA, which is based on the simulated blood flow in the LA and LAA, the derived hemodynamic parameters being one or more elements from the relative residence time (RRT), energy loss, pressure loss coefficient, wall shear stress (WSS) or oscillation index (OSI).
  • Additionally or alternatively, the means for calculating hemodynamic features for the LA and LAA on the basis of the simulated blood flow in the LA and LAA can comprise:
    • Means for calculating, for at least several locations in the LAA, a statistical identification of at least one element from blood flow velocity or pressure, on the basis of the simulated blood flow in the LA and LAA.
    • In addition or as an alternative, the means for calculating the derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the cardiac function on the basis of the measurements of the LA and
    • LAA extracted from the patient's medical image data also include:
      • Means for simulating the electrical signal propagation in the LA using the patient-specific calculation model of the heart function; and
      • Means for calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being input as features into the classifier which is based on the machine learning.

According to a further preferred development of the invention, an apparatus can be provided, wherein the extracted LA and LAA measurements, which are input as features into the classifier, which is based on machine learning, the LA volume, LAA volume and the number the LAA flap, and the calculated derived metrics that are entered as features into the classifier based on machine learning include one or more items from Relative Residence Time (RRT), Energy Loss, Pressure Loss Coefficient, Wall Shear Stress (WSS), Oscillation index (OSI), mean blood flow velocity or mean blood pressure, at one or more points in the LA and LAA.

• Extrahieren von Messungen am linken Atrium (LA) und linken Atriumsanhang (LAA) aus medizinischen Bilddaten eines Patienten;
• Berechnen von abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines patientenspezifischen Berechnungsmodells der Herzfunktion auf der Basis der Messungen des LA und LAA, die aus den medizinischen Bilddaten des Patienten extrahiert wurden; und Berechnen einer Schlaganfallrisikoeinstufung für den Patienten auf der Basis der extrahierten LA- und LAA-Messungen und der berechneten abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines Klassifikators, der auf dem trainierten maschinellen Lernen basiert, wobei die extrahierten LA- und LAA-Messungen und die berechneten abgeleiteten Metriken für das LA und LAA als Merkmale in den Klassifikator eingegeben werden, der auf dem trainierten maschinellen Lernen basiert.
• Nichtflüchtiges computerlesbares Medium nach Anspruch 23, wobei das Extrahieren von Messungen des linken Atriums (LA) und des linken Atriumsanhangs aus medizinischen Bilddaten eines Patienten folgendes umfassen kann:
  • Segmentieren des LA in den medizinischen Bilddaten des Patienten.

In order to overcome the disadvantages, the invention proposes, according to a second aspect, a non-transitory computer-readable medium which stores computer program instructions for patient-specific stroke risk prediction on the basis of medical images, the computer program instructions, when executed by a processor, causing the processor Performs operations that include:

• Extracting measurements on the left atrium (LA) and left atrium appendage (LAA) from medical image data of a patient;
• Computing derived metrics for the LA and LAA of the patient using a patient-specific computation model of the cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data; and calculating a stroke risk score for the patient based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a classifier based on the trained machine learning, the extracted LA and LAA Measurements and the calculated derived metrics for the LA and LAA are entered as features into the classifier, which is based on the trained machine learning.
• The non-transitory computer readable medium of claim 23, wherein extracting left atrial (LA) and left atrial appendage measurements from medical image data of a patient may include:
  • Segmenting the LA in the patient's medical image data.

• Extrahieren eines LA-Volumens, LAA-Volumens und einer Anzahl von LAA-Lappen auf der Basis des segmentierten LA, wobei das LA-Volumen, LAA-Volumen und eine Anzahl von LAA-Lappen als Merkmale in den auf maschinelles Lernen basierten Klassifikators eingegeben werden.

According to a further preferred development of the invention, a non-transitory computer-readable medium can be provided, wherein the extraction of the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient further comprises:

• Extracting an LA volume, LAA volume and a number of LAA lobes based on the segmented LA, with the LA volume, LAA volume and a number of LAA lobes being entered as features into the machine learning based classifier .

• Extrahieren von hämodynamischen Messungen für das LA und LAA aus den medizinischen Bilddaten des Patienten.

According to a further preferred development of the invention, a non-transitory computer-readable medium can be provided, wherein the extraction of the measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient further comprises:

• Extract hemodynamic measurements for the LA and LAA from the patient's medical image data.

• Simulieren der Blutströmung im LA und LAA unter Verwendung des patientenspezifischen Berechnungsmodells der Herzfunktion; und
• Berechnungen von thermodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA, wobei die hämodynamischen Merkmale in den Klassifikator, der auf dem trainierten maschinellen Lernen basiert, eingegeben werden.
• Das Berechnen von hämodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA kann umfassen:
  • Berechnen von abgeleiteten hämodynamischen Parametern für mindestens mehrere Orte im LAA, das auf dem simulierten Blutstrom im LA und LAA basiert, wobei die abgeleiteten hämodynamischen Parameter ein oder mehrere Elemente aus der relativen Verweilzeit (RRT), Energieverlust, Druckverlustkoeffizient,
  • Wandschubspannung (WSS) oder Oszillationsindex (OSI) enthalten.
  • Das Berechnen von hämodynamischen Merkmalen für das LA und LAA auf der Basis des simulierten Blutstroms im LA und LAA umfasst:
• Das Berechnen für mindestens mehrere Orte im LAA, eine statistische Kennzeichnung von mindestens einem Element aus Blutströmungsgeschwindigkeit oder Druck, kann auf dem simulierten Blutstrom im LA und LAA basieren.
• Außerdem oder alternativ kann das Berechnen der abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines patientenspezifischen Berechnungsmodells der Herzfunktion auf der Basis der Messungen des LA und LAA, die aus den medizinischen Bilddaten des Patienten extrahiert wurden, ferner umfassen:
  • Simulieren der elektrischen Signalausbreitung im LA unter Verwendung des patientenspezifischen Berechnungsmodells der Herzfunktion; und
  • Berechnen von elektrophysiologischen Merkmalen für das LA und LAA aus der simulierten elektrischen Signalausbreitung im LA, wobei die berechneten elektrophysiologischen Merkmale als Merkmale in den Klassifikator eingegeben werden, der auf dem maschinellen Lernen basiert.

According to a further preferred development of the invention, a non-transitory computer-readable medium can be provided, wherein the calculation of derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the heart function on the basis of the LA and LAA measurements obtained from the medical Image data extracted from the patient includes:

• Simulating the blood flow in the LA and LAA using the patient-specific computation model of the heart function; and
• Calculations of thermodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being entered into the classifier, which is based on the trained machine learning.
• Computing hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA can include:
  • Calculation of derived hemodynamic parameters for at least several locations in the LAA, which is based on the simulated blood flow in the LA and LAA, the derived hemodynamic parameters being one or more elements from the relative residence time (RRT), energy loss, pressure loss coefficient,
  • Wall shear stress (WSS) or oscillation index (OSI) included.
  • Computing hemodynamic characteristics for the LA and LAA based on the simulated blood flow in the LA and LAA includes:
• The calculation for at least several locations in the LAA, a statistical identification of at least one element from blood flow velocity or pressure, can be based on the simulated blood flow in the LA and LAA.
• Additionally or alternatively, calculating the derived metrics for the patient's LA and LAA using a patient-specific calculation model of the cardiac function based on the measurements of the LA and LAA extracted from the patient's medical image data may further include:
  • Simulating the electrical signal propagation in the LA using the patient-specific calculation model of the heart function; and
  • Calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being entered as features into the classifier, which is based on machine learning.

According to a further preferred development of the invention, a non-transitory computer readable medium can be provided, wherein the extracted LA and LAA measurements, which are input as features into the classifier, which is based on machine learning, LA volume, LAA volume and the Include number of LAA lobes, and the calculated derived metrics entered as features into the classifier based on machine learning include one or more items from Relative Residence Time (RRT), Energy Loss, Pressure Loss Coefficient, Wall Shear Stress (WSS) , Oscillation Index (OSI), mean blood flow velocity or mean blood pressure, at one or more points in the LA and LAA.

• Berechnen der Schlaganfallrisikoeinstufung des Patienten auf der Basis der extrahierten LA- und LAA-Messungen und der berechneten abgeleiteten Metriken für das LA und LAA des Patienten unter Verwendung eines trainierten tiefen neuronalen Gitters.

According to a further preferred development of the invention, a non-transitory computer readable medium can be provided, wherein the calculation of a stroke risk rating of the patient based on the extracted LA and LAA measurements and the calculated derived metrics for the LA and LAA of the patient, wherein a classifier which is based on trained machine learning includes:

• Calculating the patient's stroke risk score based on the extracted LA and LAA measurements and the calculated derived metrics for the patient's LA and LAA using a trained deep neural grid.

These and other advantages of the invention will become apparent to those skilled in the art by reference to the following detailed descriptions and accompanying drawings.

Figure list

• 1 illustrates a method for patient-specific ischemic stroke risk prediction based on medical images according to an embodiment of the present invention;
• 2 illustrates an exemplary triangular left atrium (LA) grid segmented from CT data;
• 3 illustrates a partially based model of the left atrium (LA) according to an embodiment of the present invention;
• 4th illustrates an example of simulated blood flow in the LA;
• 5 illustrates an exemplary feature map depicting a hemodynamic feature field on the surface of the LA; and
• 6th Figure 3 is a high-level block diagram of a computer capable of implementing the present invention.

DETAILED DESCRIPTION

The present invention relates to a patient-specific ischemic stroke risk prediction based on an automatic analysis of the left atrium (LA) and the left atrial appendage (LAA) in medical images. A digital image is often composed of digital representations of one or more objects (or shapes). The digital representation of an object is often described herein in terms of identifying and manipulating the objects. Such manipulations are virtual manipulations that are achieved in the memory or other circuits / hardware of a computer system. Accordingly, it will be understood that embodiments of the present invention may be practiced within a computer system using data stored in the computer system or available through a trellis system.

Embodiments of the present invention provide ischemic stroke risk prediction for a patient based on automated analysis of the LA and LAA. While atrial fibrillation, atrial flutter, sinus node dysfunction / asystole, arrhythmias, atrial septal aneurysms, and the Chiari trellis are all associated with left atrial thrombi, the LAA is considered to be the primary site of formation. The mechanism for such formation involves stagnation of blood in the LAA due to inefficient blood flow associated with, for example, pathological contraction patterns. Embodiments of the present invention provide patient-specific ischemic (cardioembolic) stroke risk stratification using biomarkers based on characteristics associated with the LAA as well as "exogenous" biomarkers, including type of medication. The features associated with the LAA can include morphological features, hemodynamic features, and "hidden" variables that are determined by a machine learning algorithm. Embodiments of the present invention rely on computational modeling to reveal various factors that play a role in LAA thrombus formation and the eventual risk of embolic stroke. Embodiments of the present invention utilize patient-specific anatomical and computational modeling to determine known risk factors for LAA thrombus formation, including CHADS2 score, LA volume, left ventricular ejection fraction (LVEF) and degree of dense spontaneous echo contrast (SEC), as well as novel factors of interest, including complexity of LAA morphology (e.g. number of LAA lobes) and relative residence time (RRT).

1 illustrates a method for patient-specific ischemic stroke risk prediction according to an embodiment of the present invention. In step 102 medical image data of a patient are received. The medical image data may be acquired using some type of medical imaging modality, such as magnetic resonance imaging (MRI), 3D echocardiogram (echo), three-dimensional and rotational angiography, ultrasound (US), etc., provided that the heart is in the medical image data is visible. In an advantageous implementation, the medical image data include three-dimensional (3D) and / or 4D (3D + time) medical image data. The medical image data can be received directly from an image acquisition device, such as an MRI scanner, a CT scanner, a C-arm image acquisition unit, or a US scanner, or the medical image data can be received by loading previously stored medical image data of the patient will. In addition to the medical image data, clinical measurements of the patient can also be obtained. For example, clinical data such as electrocardiogram (EKG) measurements and / or pressure cuff measurements of the patient can be obtained. EP features such as coarse fiber waves, atrial flutter and / or multifunctional atrial tachycardia (MAT) can be extracted from the patient's EKG measurements.

In step 104 LA and LAA measurements are extracted from the patient's medical image data. In particular, a patient-specific anatomy / morphology of the LA and the LAA can be extracted from the medical image data of the patient by generating a patient-specific anatomical model of the LA. In order to generate a patient-specific anatomical model of the LA, the medical image data (e.g. 3D echo, MRI, CT, DynaCT) are used to generate a segmentation of the LA, which is either explicitly represented as a grid (e.g. triangular grid ), or implicitly, for example by means of a signed or unsigned distance function or a level set. 2 illustrates an exemplary triangular grid 200 of an LA segmented from CT data. A single grid that corresponding to a given timestamp can be generated, or a sequence of grids covering part or all of the cardiac cycle can be generated, thus capturing the wall movement of the LA. In one possible implementation, only the LA is segmented in the medical image data. In other possible implementations, a more complete segmentation of the left heart (ie, LA and left ventricle) or a segmentation of the whole can be created. In addition to the patient's anatomy / morphology, other LA and LAA measurements, such as hemodynamics or electrophysiology information, can also be extracted. For example, hemodynamic information, such as blood flow rate measurements, can be extracted from Doppler echo images and / or phase contrast MRI (PC-MRI) images of the patient. Electrophysiology (EP) measurements can be extracted from a patient's EKG and by personalizing a computational EP model for the patient. For example, nerves and / or fibers in the LA can be modeled as part of the anatomical model of the LA.

In an advantageous embodiment, an anatomical model of the LA can be generated by segmenting the LA in the medical image data using a multi-part atrium model. 3 illustrates a parts-based model of the left atrium (LA) according to an embodiment of the present invention. As in picture (a) of 3 shown contains the part-based LA model 300 the LA chamber body 302 , the left atrium annex (LAA) 304 and 4 major pulmonary veins (PVs) 306 , 308 , 310 and 312 . The 4 larger PVs are the lower left PV 312 , the upper left PV 310 , the lower right PV 308 and the upper right PV 306 . The shape of the LAA 304 strongly resembles an inclined cone, and the PVs 306 , 308 , 310 and 312 each have a tubular structure. Every LA part 302 , 304 , 306 , 308 , 310 and 312 is a much simpler anatomical structure compared to a holistic LA structure and therefore can be identified and segmented using a model-based approach. According to an advantageous embodiment, the LA chamber body 302 and LAA 304 are first segmented together using a Marginal Space Learning (MSL) framework. The idea of MSL is not to teach in a monolithic classifier directly in the full similarity transformation parameter space, but to teach in classifiers step by step at marginal spaces. In particular, the determination of each ventricle can be broken down into three problems. Position estimation, position-orientation estimation and position-orientation scale estimation. A separate classifier is trained on the basis of annotated training data for each of these estimation problems. Each classifier can be a classifier with a probabilistic reinforcement tree (PBT) that is trained on the basis of annotated training data. The classifiers in the lower dimensional edge spaces are used to efficiently reduce the search space. This object localization stage results in an estimated transformation (position, orientation and scale) of the object (for example the heart chamber). After the automatic object localization, the mean shape model of the object is aligned with the estimated transformation in order to obtain a rough estimate of the object shape. The shape is then locally deformed to adapt the object boundary. Active Shape Models (ASM) can be used to deform an initial estimate of a non-rigid shape under the guidance of the image credit and the previous shape. However, a non-learning exemplary boundary detector such as that used in conventional ASM applications does not work effectively in deforming the heart chamber due to the complex background and faint edges. Instead, a learning boundary detector can be used to take advantage of more image evidence to obtain robust boundary detection. Additional details regarding MSL-based ventricular segmentation are provided in US 7 916 919 B2 , which was issued on March 29, 2011 and is entitled "System and Method for Segmenting Chambers of a Heart in a Three Dimensional Image". US 2010/0 040 272 A1 and US 2012/0 022 843 A1 which are incorporated herein by reference. Once the LA chamber body 302 and LAA 304 are segmented, the remaining LA parts become 306 , 308 , 310 and 312 segmented using MSL-based segmentation, depending on a statistical shape constraint on the segmented LA chamber body 302 based.

Once the LA parts are segmented in the medical image data, they are combined into a consolidated grid model. Image (b) of 3 shows a consolidated LA grid 320 who have favourited the LA Chamber 322 , LAA 324 and PVs 326 , 328 , 330 and 332 includes. Additional details regarding the methods for part-based atrial segmentation are provided in US 8 644 576 B2 and US 8 724 881 B2 which are incorporated herein by reference in their entirety. In one possible implementation, the right atrium (RA) can also be similarly segmented, either holistically or using a parts-based method using MSL-based segmentation, and the LA and RA models can be combined into a two-atrium grid. Atrial anatomical structures, such as the SA node, Bachmann bundles (BB), crista terminalis and comb-shaped muscles (PM) can be identified on the lattice apices of the atrial lattice.

A fiber orientation for the LA can be modeled based on historical observations. In one possible implementation, it can be assumed that the atrial tissue is isotropic and that it is equally likely to spread in all directions. Alternatively, if in vivo diffusion tensor (DT) MR images are available, DT MR images of the patient's cardiac fibers can be mapped directly onto the anatomical model through image registration. In this case, the DT-MR image is recorded non-linearly on the medical image in which the LA model is segmented. The resulting transformation is used to deform the tensor field in the DT-MR image in the direction of the anatomical model. The finite strain model, the details of which are described in Peyrat et al., "A Computational Framework for the Statistical Analysis of Cardiac Diffusion Tensors: Application to a Small Database of Canine Hearts", IEEE TMI, 26 (11): 1500-1514, 2007 , which is incorporated herein by reference, can be used to reorient the tensors once the tensors are registered in the anatomical model. It is also possible that an atlas of the fiber architecture is available and the atlas is registered for the patient-specific LA anatomical model using standard image registration techniques.

The local atrial wall thickness is non-uniform with values between 0.8 mm and 3 mm. The regional atrial wall thickness can be extracted from high resolution MRI images, but it can be time consuming to extract the atrial wall thickness for the entire atrium. An alternative way to generate the atrial model with non-uniform wall thickness is to make a grid thickening (for the non-ablated areas) using level set thresholding from the patient's atrial images.

If tissue fibrosis (for example scar tissue) can be identified in the medical images, this information is also incorporated into the patient-specific anatomical LA model. For example, scar tissue and boundary tissue can be segmented using DE-MRI image data. The anatomical 3D model of the LA can be strictly registered on a DE-MRI image, using the coordinates of the MR scanner plus correlations between image information in the DE-MRI image and the anatomical 3D image. An anticipation minimization algorithm with prior conviction and spatial regularization can then be used to segment the scar and boundary tissue. This method works on in vivo multimodal images and adds sliding constraints for increased robustness against noise. Healthy tissue and scar tissue are modeled using a mixed Gaussian model with two modes. If a three-class segmentation is given, the parameters of the mixed model are estimated, from which a conviction value λ is derived. Voxels (volume elements) with λ <0.5 are rejected by the model and classified as a border zone. The border zone is a zone surrounding the scar tissue, which represents healing tissue. For increased robustness and regularity, Markov random fields are used to reject voxels according to the state of neighboring voxels. Furthermore, voxels further than N mm from the current scar estimate are never rejected, assuming that the boundary zone can only be found in the vicinity of scars. A graph cut algorithm can be used to estimate smooth interfaces between tissue types. The graph cut algorithm is initiated with a rough classification obtained using a k-mean algorithm or a similar algorithm, and it is repeated until convergence (e.g. when the parameters of the blended model no longer change). The segmented scar tissue and the surrounding border zone are then mapped onto the volumetric grid representation of the LA.

One turns to 1 back, be in step 106 the derived metrics for the LA and the LA and LAA are calculated using a patient-specific calculation model of the cardiac function on the basis of the LA and LAA measurements made in step 104 were extracted. The computational model of the cardiac function may include modeling the wall mechanics and electrical signal propagation in a fully coupled or decoupled manner. In an advantageous embodiment, which is described in more detail below, a patient-specific EP calculation model can use the Lattice-Boltzmann method for electrophysiology (LBM-EP) in order to simulate the electrical signal propagation in the LA. This can be done using the whole heart, only the left heart (i.e. LA and left ventricle), only the atria (i.e. LA and right atria), or just the LA. If the anatomical model that is in step 104 less than the whole heart, models with reduced dimensionality can be used for the missing anatomical components to enable the EP calculation. Pathological EP conditions, such as atrial fibrillation, can be included in the patient-specific calculation model and can be used to generate the associated pathological wall kinemathek, which is used in the stroke risk classification from step 108 can be used. In addition, simulated EP values such as electrical activation time and duration of the action potential can be calculated for each point in the LA and LAA using the patient-specific EP calculation model, and derived EP parameters can be calculated based on the simulated EP values . Exemplary inferred EP parameters can include chaotic polarization patterns and the permanent presence of rotors. In one possible implementation, the simulated EP values can be used to generate a simulated EKG signal for the patient, and inferred features such as coarse fiber waves, atrial flutter and / or multivocal atrial tachycardia (MAT) can be derived from the simulated EKG Signal can be extracted. The simulated EP values and / or the derived EP parameters can be used as features for machine learning on the basis of the stroke risk classification in step 108 can be entered.

The patient-specific EP calculation model is a calculation model of LA electrophysiology that is personalized by estimating the patient-specific parameters of the EP calculation model that represents tissue properties of the LA tissue based on the measured EP data of the patient. The patient-specific EP calculation model of the LA simulates the electrical signal propagation in the LA. According to an advantageous implementation, the EP calculation model uses the Lattice-Boltzmann method for electrophysiology (LBM-EP) in order to solve a monodomain mountain model of the patient-specific LA anatomy using a multicellular model. In this method, a characteristic grid domain for electrophysiology calculations is calculated using the patient-specific anatomical model of the LA. A Cartesian grid, with uniform grid spacing or possibly with unequal and spatially varying spacing, is first created in a bounding rectangle that surrounds the anatomical model of the LA. The grid spacing can be defined by the user or set in the system. A level-adjusted representation is then calculated from the patient-specific anatomical model as follows. For each node x of the grid, the shortest distance to the anatomical model grid is calculated and assigned to this node. In an advantageous embodiment, nodes within the myocardium are defined by positive distances, nodes that are not in the myocardium are defined by negative distances. The opposite convention can be used well without modification. Nodes on the myocardium, endocardium, and picardium are identified as such, as are other nodes belonging to the anatomical atrial structures that are important in directing atrial excitation. For example, the nodes of the patient-specific anatomical model can be marked on the categorical grid corresponding to SA, BB, crista terminalis, crest muscles and appendages. Available scars and border areas are also reported in the domain by additional level set information, and the conductivity for such regions can be set to a predetermined reduced value or to zero. The fiber orientation f (x) is mapped onto each node using screening methods, or it is recalculated directly from the associated endocardial and epicardial zones. Cell model parameters, for example ionic conductivity c (x), are assigned to each node x.

wobei R(x, t) ein Reaktionsterm ist, der die zellulären Mechanismen beschreibt, welche das Aktionspotenzial hervorrufen, c(x) ist die lokale lonenstromleitfähigkeit, D(x) ist die Anisotropie-(transversale Isotropie)-Matrix, die durch (1-ρ)f(x)f(x)^T + pld definiert ist, wobei p das Verhältnis zwischen der Kreuzfaserdiffusivität und der Faserdiffusivität ist (normalerweise ρ = 0,11 - 0,25). Es ist auch möglich, vollständig isotrope Tensoren D(x) für die verbesserte Charakterisierung der Vorhoffaserarchitektur zu verwenden.The LA EP calculation model calculates the transmembrane potential at each node within the LA using the Lattice-Boltzmann Method for Electrophysiology (LBM-EP). The EP calculation model calculates the variation of the transmembrane potential v (x, t) over time according to the monodomain equation: $$\frac{dv\left(x,t\right)}{dt}=\mathrm{R.}\left(x,t\right)+\nabla \cdot c\left(x\right)\mathrm{D.}\left(x\right)\nabla \text{v}\left(x,t\right),$$

where R (x, t) is a reaction term that describes the cellular mechanisms that cause the action potential, c (x) is the local ionic conductivity, D (x) is the anisotropy (transverse isotropy) matrix, which is defined by (1 -ρ) f (x) f (x) ^T + pld, where p is the ratio between the cross fiber diffusivity and the fiber diffusivity (usually ρ = 0.11-0.25). It is also possible to use fully isotropic tensors D (x) for the improved characterization of the atrial fiber architecture.

The choice of the response term R (x, i) depends on the cellular model of cardiac electrophysiology that is used. According to an advantageous embodiment of the present invention, a multicellular EP model can be used to model the EP of the atria. In order to describe the effect of non-homogeneity of the tissue of the atrial EP, the Cartesian nodes in the calculation domain of the LBM-EP solver belonging to the different anatomical atrial structures can be identified and fixed with different cellular models and / or electrical conductivity values. In an exemplary implementation, the Courtemanche-Ramirez-Nattel (CRN) human atrial cell model described in Courtemanche et al. is described, "lonic Mechanisms Underlying Human Atrial Action Potential Properties: Insights from a Mathematical Model", Am. J. Physiol. 275, H301-H321 (1998) , can be used as atrial cell model. The CRN atrial cell model has 35 static parameters and 21 ordinary differential equations to describe 12 ion channels, the corresponding gating variables and ion concentrations.

Equation (1) is solved using the Lattice-Boltzmann method for electrophysiology, referred to herein as LBM-EP. LBM-EP is a highly parallelizable algorithm for solving monodomain electrophysiology equations. The LBM-EP algorithm is detailed in the US 2013/0 226 542 A1 , entitled "Method and System for Fast Patient-Specific Cardiac Electrophysiology Simulations for Therapy Planning and Guidance," which is incorporated herein by reference in its entirety. In contrast to the standard finite element method, LBM-EP does not explicitly solve the reaction-diffusion equation, but rather calculates the "movement" of particles on a Cartesian grid, from which the reaction-diffusion behavior results. The particles can move according to fixed directions (or connectivities) with a certain probability. The algorithm comprises two nodal steps: streaming, which causes the particle to jump from one node to another; and Collision, which takes care of mass conservation and boundary conditions. It can be shown mathematically that this simple algorithm reproduces the dynamics of the reaction-diffusion equation. To calculate cardiac electrophysiology using LBM-EP, domain boundaries are represented as level sets and tissue anisotropy is modeled. In the sinus rhythm, the electrocardiography model can be calculated with periodic stimulus on the septum in order to imitate the fast-conducting bundle of His.

Since the LBM-EP method is completely oriented towards nodes and the time integration is explicit, the calculations can be carried out locally, and the method can therefore easily be adapted to strongly parallel architectures. In an advantageous embodiment, the method can be implemented in one or more general purpose graphics processing units (GPGPU), which enables an almost real-time and accurate cardiac electrophysiology calculation during the procedure. In this embodiment, this method can be optimized to take full advantage of the computing power of GPGPUs. For example, adaptive computation block aggregation can be performed to strike a balance between computation performance and memory bandwidth. Adaptive time step methods can also be implemented to account for the current EP dynamics, especially in sinus rhythm regions. For example, small time steps can be used when the fast front propagation is occurring, larger time steps can be used during the refractory period, and even larger time steps can be used during the depolarization state. Adaptive lattice refinement methods can also be implemented in order to reduce the overall computational effort in the presence of thick-walled structures (e.g. ventricles) and thin-walled structures (e.g. atria). The model can be linked to model memories for greater flexibility.

wobei λ ein konstantes Diffusions-Anisotropie-Verhältnis ist, λ = D_i(x)/D_e(x), und D_i und D_e sind intra- bzw. extrazelluläre Tensoren für das Diffusionsvermögen. Das extrazelluläre Potenzial Φ_e wird dann zurück auf das Vorhofflächengitter unter Verwendung der trilinearen Interpolation abgebildet. Die extrazellulären Potenziale werden dann auf ein Torso-Oberflächen-Gitter unter Verwendung eines Grenzelementverfahrens (BEM) projiziert. Das Potenzial Φ(x) an einem beliebigen Punkt x der Thoraxdomäne (Torso-Oberflächen-Gitter) kann folgendermaßen berechnet werden: $$\varphi \left(x\right)=\frac{1}{4\pi }{\displaystyle {\int }_{S_B}{\varphi }_b}\frac{r\cdot n}{{\Vert r\Vert }^3}d{S}_B-\frac{1}{4\pi }{\displaystyle {\int }_{S_H}\left[{\varphi }_e\frac{r\cdot n}{{\Vert r\Vert }^3}+\frac{\nabla {\varphi }_e\cdot n}{\Vert r\Vert }\right]}d{S}_H,$$

wobei r der Vektor ist, der durch x und den Integrationspunkt n definiert ist, während S_B und S_H die Torso- bzw. die Epikardfläche sind. Das Körperflächenpotential am Torso, Φ_b, kann als Funktion des extrazellulären Potentials Φ_e ausgedrückt werden, welches ermöglicht, dass das Potential an einem beliebigen Punkt auf dem Torso berechnet werden kann. Das Torsogitter kann aus den Daten des medizinischen Bildes segmentiert werden, wobei Maschinenlernalgorithmen verwendet werden. Entsprechend einer möglichen Implementierung kann das Körperflächenpotential Φ_b für jeden Scheitelpunkt auf dem Torsogitter berechnet werden. In einer weiteren möglichen Implementierung kann das Körperflächenpotential Φ_b nur für Scheitelpunkte auf dem Torsogitter berechnet werden, die den Orten von Leitungen entsprechen, die zum Erfassen der gemessenen EKG-Signale des Patienten verwendet werden (z.B. 12-Leitungs-EKG). Ein simuliertes EKG-Signal wird unter Verwendung der Körperflächenpotentiale, die an den Positionen der EKG-Leitungen berechnet wurden, berechnet, und die EKG-Merkmale, wie zum Beispiel die Dauer des QRS-Komplexes Δ_QRS und der elektrische Achsenwinkel α_EA, können automatisch aus dem simulierten EKG-Signal abgeleitet werden. Es sollte beachtet werden, dass in der obigen Beschreibung ein homogenes Torsomodell eingesetzt wird. Dieses kann jedoch auf ein heterogenes Torsomodell erweitert werden, das Muskeln, Lunge, Knochen, Fett und andere Gewebe enthält, wie sie in medizinischen Bildern identifiziert werden. Jedes Gewebe würde dann eine andere elektrische Leitfähigkeit haben.The EP calculation model of the LA can be coupled with a boundary element model of the potential propagation in soft tissue in order to calculate an EKG resulting from the simulated cardiac electrophysiology. This enables body area EKG measurements of the patient to be transferred back to the atrial model for personalizing the EP calculation model. The EP computation model of the atria computes a transmembrane potential for each node of the patient-specific anatomical model in the computation domain at each time step. An extracellular potential Φ _e is calculated at each node of the computational domain based on the transmembrane potential v (x, t) using a closed-form expression (Ω defines the computational domain; | Ω | is the number of elements in it): $${\varphi }_e\left(\text{x},\text{t}\right)=\frac{\lambda }{1+\lambda }\frac{1}{\left|\mathrm{\Omega }\right|}{\displaystyle {\int }_{\mathrm{\Omega }}\left[v\left(,y,t\right)-v\left(x,t\right)\right]}dy,$$

where λ is a constant diffusion-anisotropy ratio, λ = D _i (x) / D _e (x), and D _i and D _e are intra- and extracellular tensors for the diffusivity, respectively. The extracellular potential Φ _e is then mapped back onto the atrial surface grid using trilinear interpolation. The extracellular potentials are then projected onto a torso surface grid using a boundary element method (BEM). The potential Φ (x) at any point x in the thorax domain (torso surface grid) can be calculated as follows: $$\varphi \left(x\right)=\frac{1}{\mathrm{4th}\pi }{\displaystyle {\int }_{{\mathrm{S.}}_{\mathrm{B.}}}{\varphi }_b}\frac{r\cdot n}{{\Vert r\Vert }^3}d{\mathrm{S.}}_{\mathrm{B.}}-\frac{1}{\mathrm{4th}\pi }{\displaystyle {\int }_{{\mathrm{S.}}_H}\left[{\varphi }_e\frac{r\cdot n}{{\Vert r\Vert }^3}+\frac{\nabla {\varphi }_e\cdot n}{\Vert r\Vert }\right]}d{\mathrm{S.}}_H,$$

where r is the vector defined by x and the integration point n, while S _B and S _{H are} the torso and epicardial surfaces, respectively. The body area potential on the torso, Φ _b , can be expressed as a function of the extracellular potential Φ _e , which enables the potential at a any point on the torso can be calculated. The torso grid can be segmented from the medical image data using machine learning algorithms. According to one possible implementation, the area potential Φ _b can be calculated for each vertex on the torso lattice. In a further possible implementation, the body _{surface potential Φ b} can only be calculated for vertices on the torso grid that correspond to the locations of lines that are used to acquire the measured EKG signals of the patient (eg 12-lead EKG). A simulated EKG signal is calculated using the body area potentials calculated at the positions of the EKG leads, and the EKG features, such as the duration of the QRS complex Δ _QRS and the electrical axis angle α _EA , can automatically can be derived from the simulated ECG signal. It should be noted that a homogeneous torso model is used in the above description. However, this can be extended to a heterogeneous torso model that includes muscles, lungs, bones, fat, and other tissues as identified in medical images. Each tissue would then have a different electrical conductivity.

The LA's EP calculation model must be personalized in order to be prognostic for a specific patient. The EP calculation model of the LA is personalized on the basis of EP measurements of the patient, such as invasive cardiac EP maps or body surface potential maps. In an exemplary embodiment, electrophysiology (EP) data of the patient is first merged with the patient-specific anatomical model of the atria. The patient's EP data may include invasive cardiac EP maps that are captured for the patient and / or body surface potential measurements. Body surface potential measurements are measurements of electrical potentials on the patient's torso and can refer to body surface potential maps that were recorded using body surface mapping (BSM) or EKG measurements of the patient using EKG leads on the patient's torso ( e.g. 12-lead EKG measurements). Invasive cardiac maps are generated invasively by measuring potentials at various points in the heart over time, for example using a catheter mapping system or a catheter basket system. In order to personalize the EP calculation model using invasive cardiac EP maps, the cardiac EP data are registered for the patient-specific anatomical model of the atria.

In order to combine the body surface potential measurements (e.g. body surface potential maps obtained using BSM or EKG measurements) with the patient-specific anatomical model of the atria, the body surface measurements are mapped onto a patient-specific torso model that is used for the patient-specific anatomical model of the atria is registered. A 3D image of the patient's torso can be captured, e.g., at the start of the intervention, and a triangulated grid of the patient's torso can be segmented from the 3D image using a segmentation algorithm such as graph cuts. In a case where 3D torso images cannot be captured, 2D MRI scout images can be used to create a torso model. In this case, the contours of the visible torso in the 2D scout images can be extracted automatically, e.g. using graph cuts. A stored 3D torso atlas can then be registered based on the 2D scout images using an affine transformation to fit the torso contours extracted from the 2D scout images. This registration algorithm supports the scout image positions (axial, sagittal and coronal) for increased robustness and minimizes the dangers of local minima. Once the torso is modeled, the patient-specific anatomical model can be automatically registered for the heart model using the scanner coordinates. Line positions from which the body surface potentials were measured are mapped onto the torso model. For example, the lead positions can be mapped automatically or using custom landmarks. The body surface potential measurements can then be back-projected onto the patient-specific anatomical model of the torso.

Once the patient-specific torso model is registered with the patient-specific anatomical model of the atria, an electrical model of the diffusion in the torso can be used to describe the coupling relationship between the heart and the torso. As described above, electrical potentials on the torso can be calculated from cardiac potentials by first inferring extracellular potentials from transmembrane potentials on the atrium and then solving Poisson's equation using the Boundary Element Method (BEM). The electrical coupling between the heart grid and the torso grid can be _{modeled by the linear relationship ∀ t} , Y _t = T * X _t , where X _t (x) denotes the extracellular potentials on the epicardium, Y _t (x) denotes the torso potentials and T is the coupling matrix created by boundary element discretization the heart-torso geometry and solving Poisson's equation for electrical potentials.

The patient's EP measurements are used to personalize the electrical conductivity values and multicell model parameters at the nodes of the computational domain. For a detailed multicell model that is specific to the atria, such as the CRN atrial cell model, which has 35 parameters, a model reduction approach can be used to reduce the computation costs and the number of model parameters. For example, a reduced model can be constructed by action potential (AP) manifold learning to reduce the number of model parameters, followed by learning a regression model to predict the parameters in the reduced AP manifold. This speeds up the personalization process, because the more parameters to be estimated, the greater the number of iterations that are required for the forward solver.

In one embodiment, personalization of the EP computation model may be performed using an inverse problem approach based on a comparison of the simulated EP data generated by performing simulations using the EP computation model and the patient's EP measurements . In particular, personalization can be formulated as an optimization problem with the aim of determining the point-wise difference between the calculated (simulated) activation times when using the EP calculation model and the activation times in the measured EP data of the patient over all nodes of the calculation domain (i.e. over all nodes of the patient-specific anatomical model of the atria). A coarse-to-fine strategy can be used for an increased convergence of the estimation and robustness over local minima. In the coarse-to-fine strategy, a parameter value (for example electrical conductivity c) is first estimated for the entire LA. Areas with larger errors are then partitioned and one parameter value per partition is estimated, initialized to the value of the previous step. Then the process is repeated until convergence is achieved.

In a further embodiment, machine learning methods can be used to estimate personalized parameters of the EP calculation model of the LA on the basis of the EP measurements of the patient. In an offline training stage, a large database of activation maps or other EP measurements is created using the EP calculation model with different parameter values. Advanced nonlinear manifold learning methods are used to train a regression function and the personalized parameters are estimated online by applying the trained regression function to reverse engineer the local values of electrical conductivity and multicell model parameters when the local activation maps of the measured EP data of the Patients are given.

The computational model of the heart function can also include the modeling of the blood flow. Modeling / simulating the blood flow in the LA can be carried out using Computational Fluid Dynamics (CFD) within a domain, the boundary of which is given by the segmented LA grid. The computation domain can be static or dynamic. For a static domain, zero speed boundary conditions (which corresponds to lack of slide) can be imposed on the LA wall. For a dynamic domain, the connection between the LA wall and the fluid domain (bloodstream) can be implemented using the Fluid Structure Interaction (FSI) framework. The FSI coupling between the LA wall and the blood flow can be carried out using unilateral FSI, with the movement of the wall in advance (e.g. based on the observed wall movement in medical 4D (3D + time) image data) for each time step and the non-slip speed limit conditions force the fluid (blood) to move at the predetermined wall speed. Such a procedure is described in greater detail in the US 8 682 626 B2 , entitled "Method and System for Comprehensive Patient-Specific Modeling of the Heart," which is incorporated herein by reference in its entirety. Alternatively, a mutual FSI formulation can be used in which the wall is modeled as a deformable material, provided with material properties that correspond to the atrial tissue, and force and / or speed information on the wall is exchanged between the tissue and the blood. An exemplary way of implementing such a formulation is to use liquid pressure to create traction force boundary conditions for the deformable solid and to use the position and kinematics of the solid wall as boundary conditions for the fluid (blood). The CFD can simulate blood flow in the LA using standard Navier-Stokes equations for moment and conservation of mass, or from alternative formulations such as the Lattice-Boltzmann method (LBM). 4th illustrates an example of simulated blood flow in the LA. In particular shows 4th Blood velocity vectors 402 obtained from CFD calculations in LA. Reduced circulation within the LAA 404 can be observed in the simulated blood stream flowing in 4th will be shown.

According to an advantageous embodiment of the present invention, the patient-specific calculation model of the heart function is used to carry out a generative analysis (ie to generate features) of the hemodynamics that is specific for LA and in particular the LAA. This can contain the statistical characterization (eg mean value) of the primary hemodynamic variables, such as blood velocity and pressure and / or derived parameters specifically for the LA and / or LAA. Derived parameters refer to parameters or indices that were calculated from the calculated / simulated hemodynamics. Examples of derived parameters / indices can relate directly to hemodynamic indices, energy or pressure indices, or specific LAA flow indices. Hemodynamic indices can include indices based on wall shear stresses (WSS) in the LA or, in a possible implementation, only in the LAA (e.g. time-averaged, spatial mean (TASMWSS), time-averaged, spatially maximum WSS, time-averaged, spatial 99. Percentile WSS), low wall shear area (LSA), low shear concentration index (LSCI), low shear index (LSI), high shear area (HSA), SCI (= TASMWSS in HSA / TASMWSS in LAA), medium, maximum and 99th percentile WSS spatial gradient (WSSG), and / or mean oscillatory shear index (OSI). Energy and pressure related indices can include the energy loss in the LAA, pressure loss coefficients in the LAA, the kinetic energy ratio (KER) and / or viscous dissipation ratio (VDR). Other more specific LAA flow indices may include relative residence time (RRT), flow concentration index (ICI), and / or vortex length (VL). The statistical characterization of the primary hemodynamics parameters (e.g. mean speed and / or mean pressure) and the derived parameters can be calculated for each point of the segmented LA grid on the basis of a hemodynamics simulation over a period of time (e.g. one cardiac cycle). Such hemodynamic features can then be mapped over the area of the segmented LA, as in FIG 5 shown. 5 illustrates an exemplary feature map 500 , in which a hemodynamic feature is visibly displayed on the surface of the LA. Such a feature field can be reinforced on the LAA. EP characteristics, such as activation time and action potential duration, calculated using the EP calculation model, can also be calculated for each point of the segmented LA grid.

One returns to 1 back to step 108 , a risk score that predicts the risk of ischemic (embolic) stroke is calculated based on the measurements of the LA and LAA extracted from the medical image data and the metrics calculated using the computational model of the cardiac function. The LA and LAA measurements obtained from the medical image data in step 104 are extracted, and the LA and LAA metrics obtained using the computational model of the cardiac function in step 106 are used as features, and a machine learning method is used to calculate a stroke risk score based on these features. In an exemplary embodiment, the deep learning can be used to train a deep neural framework that learns mapping from the input features to a stroke risk score for the patient. The deep neural latticework is trained in an offline training phase based on a database of training images and fundamental truth values. For a specific patient, the measured features (e.g., morphological features) extracted from the medical image data and computational features (e.g., hemodynamic features and EP features) and calculated using the computational model of the heart function, are entered into the trained deep neural framework , and the trained deep neural trellis calculates a stroke risk rating for the patient based on the inputted characteristics. It should be understood that the present invention is not limited to deep learning, and other types of machine learning can also be used.

The measured features entered into the trained machine learning classifier may include morphological (anatomical) features including, but not limited to, LV volume, LAA volume, and the number of LAA lobes . The measured features may also include hemodynamics measurements extracted from the medical image data, such as velocity measurements and left ventricular ejection fraction (LVEF) measurements and EP measurements of the patient. The degree of dense spontaneous echo contrast (SEC) can also be measured in the medical image data and entered as a feature. The calculated features may include simulated hemodynamic features, such as statistical labels of hemodynamic parameters (e.g. blood velocity and pressure), and derived metrics, including derived hemodynamic indices, energy and pressure related indices, and other more specific LAA related flow indices, above described. Each of the hemodynamic features can be used for all grid points in the LA and for all grid points in the LAA. The calculated features can also include simulated EP features, such as electrical activation time and action potential duration for all points in the LA or all points in the LAA, as well as derived EP features, such as chaotic depolarization patterns and persistent presence of rotors, coarse fiber waves, atrial flutter and / or multifocal atrial tachycardia (MAT). Techniques based on machine learning, such as deep learning, can be used to automate the coding of "hidden" parameters associated with the feature list. In addition, conventional clinical risk factors, such as the CHADS2 score, can be calculated for the patient and entered as a feature in the classifier, which is based on the trained machine learning. Alternatively, the classifier based on the trained machine learning can calculate a risk rating without the conventional clinical risk score, and the risk score calculated by the classifier based on the trained machine learning can be calculated with the conventional risk score (for example, the CHADS2 score), which results in a composite risk score.

In an advantageous embodiment, molecular information can also be included in the list of features for calculating the stroke risk score on the basis of machine learning. For example, a feature can be added that indicates the presence of the transformed growth factor beta 1 (tgf-beta1), which according to the documents increases the vulnerability to Afib. In an advantageous embodiment, drug information can also be included in the list of features for calculating the stroke risk score on the basis of machine learning.

In step 110 the stroke risk score and representations of extracted and derived data are output. The risk score for the patient is calculated and output by the classifier, which is based on the trained machine learning, for example by displaying the risk score on a display of a computer system and storing the risk score for the patient in a memory of a computer system. In addition, one or more relevant features / metrics can be represented and displayed on a display of a computer system. For example, a 3D view of regions of stagnation (e.g., high RRT) within the LA can be created by mapping a color map corresponding to the calculated RRT values onto the extracted 3D anatomical model of the LA and displaying the LA model with the color map on a Display device can be visualized. Other features can be displayed in a similar manner.

In one embodiment of the present invention, the method of 1 can be used to plan and test the effectiveness of different treatments to see if the treatments reduce the patient's risk of stroke. For example, the method of 1 can be used to calculate an initial stroke risk score for the patient. The cardiac function computational model can then be used to simulate various Afib treatment modalities, such as pulmonary venous catheter isolation, and then recalculate the simulated cardiac function and calculated updated features from which the updated risk scores can be calculated. The updated risk scores can be compared to the initial stroke risk score to determine whether the Afib treatments that were simulated would reduce the patient's risk of stroke. Additional details regarding simulating Afib treatments using an EP computational model of the allegations are provided in US Pat US 2016/0 058 520 A1 which is incorporated herein by reference in its entirety. Similarly, the computational model can also be used to test the effect of different LAA occlusion devices or different ablation therapies on the patient's stroke risk score. The computational model can also be extended to address other sources of thrombus formation in the heart that can influence decisions relating to valve device therapy, etc., for example.

The above-described methods for patient-specific prediction of ischemic stroke risk can be implemented on a computer using known computer processors, memory units, storage devices, computer software and other components. A high-level block diagram of such a computer is shown in 6th illustrated. computer 602 contains a processor 604 that does the entire operation of the computer 602 controlled by executing computer program instructions defining such an operation. The computer program instructions can be stored in a storage device 612 (for example magnetic disk) and in the main memory 610 when execution of the computer program instructions is desired. Therefore, the steps of the method of 1 be defined by the computer program instructions stored in memory 610 and / or memory 612 are stored and used by the processor 604 which executes the computer program instructions. An image capture device 620 such as a CT scanning device, C-arm image acquisition device, MR scanning device, ultrasound device, etc., can be connected to the computer 602 can be connected to image data in the computer 602 to enter. It is possible to use the image capture device 620 and the computer 602 to implement as a device. It is also possible that the image capture device 620 and the computer 602 communicate with each other wirelessly via a lattice work. In one possible embodiment, the computer can 602 away from the image capture device 620 and can perform the procedural steps as part of a server or a cloud-based service. The computer 602 also includes one or more grid interfaces 606 to communicate with other devices via a latticework. The computer 602 also includes other input / output devices 608 that a user interaction with the computer 602 (e.g. display, keyboard, mouse, loudspeaker, etc.). Such input / output devices 608 can be used in conjunction with a set of computer programs, such as an annotation tool for annotating volumes captured by the image capture device 620 were obtained. Those skilled in the art will recognize that an actual computer implementation could include other components as well, and that 6th is a high-level illustration of some of the components of such a computer for illustrative purposes.

The medical image synthesis methods described above can be implemented using computers operating in a client-server relationship. Typically, in such a system, the client computers are remote from the server computer and influence one another via the latticework. The client-server relationship can be defined and controlled by computer programs that run on the respective client and server computers.

It is to be understood that the foregoing Detailed Description is to be understood in all respects as illustrative and exemplary, but not restrictive, and that the scope of the invention disclosed herein cannot be determined from the Detailed Description, but rather from the claims, as they are interpreted to the fullest extent permitted by the patent laws. It should be understood that the embodiments described and shown herein are only illustrative of the principles of the present invention and that various modifications can be implemented by those skilled in the art without departing from the scope and spirit of the invention. Those skilled in the art could implement various other combinations of features without departing from the scope and spirit of the invention.

Claims (23)

Method for stroke risk prediction based on medical images, comprising: Extracting measurements on the left atrium (LA) and left atrium appendage (LAA) from medical image data of a patient; Computing derived metrics for the patient's LA and LAA using a patient-specific computational model of cardiac function based on the LA and LAA measurements extracted (104) from the patient's medical image data; and Calculating a stroke risk score for the patient based on the extracted LA and LAA measurements and the calculated (106) derived metrics for the patient's LA and LAA using a classifier based on the trained machine learning, the extracted (104 ) LA and LAA measurements and the calculated (106) derived metrics for the LA and LAA are entered as features into the classifier based on the trained machine learning. Procedure according to Claim 1 wherein extracting (104) measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient comprises: segmenting the LA in the medical image data of the patient. Procedure according to Claim 2 wherein the segmenting of the LA in the medical image data of the patient comprises: generating a 3D LA grid from medical 3D image data of the patient, in particular wherein the generation of a 3D LA grid from medical 3D image data of the patient comprises: segmenting a plurality of LA parts in the medical image data using a multi-part atrium model; and creating a consolidated LA grid from the segmented multiple atrium portions. Procedure according to Claim 2 or 3 wherein segmenting the LA in the patient's medical image data comprises: generating a sequence of LA grids from 4D (3D + time) medical image data of the patient, and / or wherein extracting (104) the measurements of the left atrium (LA ) and measurements of the left atrial appendage from medical image data of a patient further comprises: extracting an LA volume, LAA volume and a number of LAA lobes based on the segmented LA, the LA volume, LAA volume and the number of LAA flaps as features are input into the machine learning-based classifier, and / or wherein extracting (104) the measurements of the left atrium (LA) and measurements of the left atrial appendage from medical image data of a patient further comprises: extracting hemodynamic measurements for the LA and LAA the patient's medical image data. Method according to one of the preceding claims, wherein the calculation of derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the cardiac function based on LA and LAA measurements extracted (104) from the medical image data of the patient , includes: Simulating the blood flow in the LA and LAA using the patient-specific computation model of the heart function; and Calculating hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being entered into the classifier which is based on the trained machine learning. Procedure according to Claim 5 wherein calculating hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA comprises: calculating derived hemodynamic parameters for at least a plurality of locations in the LAA based on the simulated blood flow in the LA and LAA, the derived hemodynamic parameters include one or more of relative residence time (RRT), energy loss, pressure loss coefficient, wall shear stress (WSS), or oscillatory index (OSI) and / or wherein calculating the hemodynamic characteristics for the LA and LAA based on the simulated blood flow in the LA and LAA based, comprises: calculating, for at least several locations in the LAA, a statistical identification of at least one element from blood flow velocity or pressure, on the basis of the simulated blood flow in the LA and LAA. Procedure according to Claim 5 or 6th wherein calculating the derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the cardiac function based on LA and LAA measurements extracted (104) from medical image data of the patient further comprises: simulating the electrical Signal propagation in the LA using the patient-specific calculation model of the heart function; and calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being input as features into the classifier based on the machine learning. A method according to any preceding claim, wherein the extracted (104) LA and LAA measurements entered as features into the machine learning based classifier are LA volume, LAA volume and the number of LAA lobes and the calculated (106) derived metrics input as features into the machine learning based classifier includes one or more items from relative residence time (RRT), energy loss, pressure loss coefficient, wall shear stress (WSS), oscillation index (OSI), mean blood flow velocity or mean blood pressure, at one or more points in the LA and LAA. The method according to any one of the preceding claims, wherein calculating a stroke risk score of the patient on the basis of the extracted (104) LA and LAA measurements and the calculated (106) derived metrics for the LA and LAA of the patient using a classifier, which is based on trained machine learning includes: Calculating the patient's stroke risk score based on the extracted (104) LA and LAA measurements and the calculated (106) inferred metrics for the patient's LA and LAA using a trained deep neural framework. Device for stroke risk prediction based on medical images, comprising: Means for extracting measurements on the left atrium (LA) and left atrial appendage (LAA) from medical image data of a patient; Means for calculating derived metrics for the patient's LA and LAA using a patient-specific calculation model of cardiac function based on the measurements of the LA and LAA extracted (104) from the patient's medical image data; and Means for calculating a stroke risk score for the patient based on the extracted (104) LA and LAA measurements and the calculated (106) derived metrics for the patient's LA and LAA using a classifier based on the trained machine learning, wherein the extracted (104) LA and LAA measurements and the calculated (106) derived metrics for the LA and LAA are entered as features into the classifier based on the trained machine learning. Device according to Claim 10 , wherein the means for extracting measurements of the left Atrium (LA) and the left atrial appendage (LAA) from medical image data of a patient comprises: means for segmenting the LA in the medical image data of the patient, in particular wherein the measurements of the left atrium (LA) and the left atrial appendage, which are shown as features in the Classifier based on the trained machine learning, include an LA volume, LAA volume and a number of LAA lobes, which are determined on the basis of the segmented LA. Device according to Claim 11 wherein the means for extracting left atrial (LA) and left atrial appendage (LAA) measurements from medical image data of a patient further comprises: means for extracting hemodynamic measurements for the LA and LAA from the patient's medical image data. Device according to one of the preceding Claims 10 until 12th wherein the means for calculating the derived metrics for the patient's LA and LAA using a patient-specific calculation model of the cardiac function based on LA and LAA measurements extracted from the patient's medical image data comprises: means for simulating the Blood flow in the LA and LAA using the patient-specific calculation model of the heart function; and means for calculating hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being input into the classifier based on the trained machine learning. Device according to Claim 13 wherein the means for calculating hemodynamic characteristics for the LA and LAA based on the simulated blood flow in the LA and LAA comprises: means for calculating derived hemodynamic parameters for at least several locations in the LAA based on the simulated blood flow in the LA and LAA based, wherein the derived hemodynamic parameters contain one or more elements from the relative residence time (RRT), energy loss, pressure loss coefficient, wall shear stress (WSS) or oscillation index (OSI), and / or wherein the means for calculating hemodynamic characteristics for the LA and LAA, which is based on the simulated blood flow in the LA and LAA, comprises: means for calculating, for at least several locations in the LAA, a statistical characterization of at least one element of blood flow velocity or pressure, on the basis of the simulated blood flow in the LA and LAA. Device according to Claim 13 or 14th wherein the means for calculating the derived metrics for the LA and LAA of the patient are extracted (104) from medical image data of the patient using a patient-specific calculation model of the cardiac function based on LA and LAA measurements, comprises: means for simulating the electrical signal propagation in the LA using the patient-specific calculation model of the heart function; and means for calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being input as features into the classifier based on the machine learning. Device according to one of the preceding Claims 10 until 15th , wherein the extracted (104) LA and LAA measurements entered as features into the classifier based on machine learning comprises one or more of the LA volume, LAA volume and the number of LAA Lobes, and the calculated (106) derived metrics input as features into the machine learning based classifier includes one or more items from Relative Residence Time (RRT), Energy Loss, Pressure Loss Coefficient, Wall Shear Stress (WSS), Oscillation index (OSI), mean blood flow velocity or mean blood pressure, at one or more points in the LA and LAA. Non-transitory computer readable medium that stores computer program instructions for patient-specific stroke risk prediction based on medical images, the computer program instructions, when executed by a processor (604), causing the processor (604) to perform operations comprising: extracting (104 ) of measurements on the left atrium (LA) and left atrium appendage (LAA) from medical image data of a patient; Computing derived metrics for the patient's LA and LAA using a patient-specific computational model of cardiac function based on the LA and LAA measurements extracted (104) from the patient's medical image data; and calculating a stroke risk score for the patient based on the extracted (104) LA and LAA measurements and the calculated (106) inferred metrics for the patient's LA and LAA using a classifier based on the trained machine learning, wherein the extracted (104) LA and LAA measurements and the calculated (106) derived metrics for the LA and LAA as features in the classifier based on the trained machine learning. Non-transitory computer readable medium after Claim 17 wherein the extraction (104) of measurements of the left atrium (LA) and the left atrial appendage from medical image data of a patient comprises: segmenting the LA in the medical image data of the patient, and / or wherein the extraction of the measurements of the left atrium (LA) and measurements of the left atrial appendage from patient medical image data further comprising: extracting an LA volume, LAA volume, and a number of LAA lobes based on the segmented LA, the LA volume, LAA volume, and the number of LAA Lobes are entered as features into the machine learning-based classifier, and / or wherein the extraction (104) of the measurements of the left atrium (LA) and measurements of the left atrial appendage from medical image data of a patient further comprises: extracting hemodynamic measurements for the LA and LAA from the patient's medical image data. Non-transitory computer readable medium according to the Claims 17 or 18th wherein calculating the derived metrics for the patient's LA and LAA using a patient-specific calculation model of cardiac function based on the measurements of the LA and LAA extracted (104) from the patient's medical image data comprises: simulating blood flow in the LA and LAA using the patient-specific calculation model of the heart function; and computing hemodynamic features for the LA and LAA based on the simulated blood flow in the LA and LAA, the hemodynamic features being input into the classifier based on the trained machine learning. Non-transitory computer readable medium after Claim 19 wherein calculating hemodynamic features for the LA and LAA on the simulated blood flow in the LA and LAA comprises: calculating derived hemodynamic parameters for at least several locations in the LAA based on the simulated blood flow in the LA and LAA, the derived hemodynamic parameters one or more of Relative Residence Time (RRT), Energy Loss, Pressure Loss Coefficient, Wall Shear Stress (WSS), or Oscillation Index (OSI), and / or wherein computing hemodynamic characteristics for the LA and LAA based on the simulated blood flow in the LA and LAA comprises: calculating, for at least several locations in the LAA, a statistical characterization of at least one element from blood flow velocity or pressure, on the basis of the simulated blood flow in the LA and LAA. Non-transitory computer readable medium according to the Claims 19 or 20th wherein calculating the derived metrics for the LA and LAA of the patient using a patient-specific calculation model of the cardiac function based on LA and LAA measurements extracted (104) from medical image data of the patient comprises: simulating the electrical signal propagation in the LA using the patient-specific calculation model of the heart function; and calculating electrophysiological features for the LA and LAA from the simulated electrical signal propagation in the LA, the calculated electrophysiological features being input as features into the classifier based on the machine learning. Non-transitory computer readable medium according to any of the preceding Claims 17 until 21 , wherein the extracted (104) LA and LAA measurements entered as features into the classifier based on machine learning include LA volume, LAA volume and the number of LAA lobes, and the calculated ( 106) derived metrics that are entered as features into the classifier based on machine learning, one or more items from relative residence time (RRT), energy loss, pressure loss coefficient, wall shear stress (WSS), oscillation index (OSI), mean blood flow velocity or includes mean blood pressure, at one or more points in the LA and LAA. Non-transitory computer readable medium according to any of the preceding Claims 17 until 22nd wherein calculating a patient's stroke risk score based on the extracted (104) LA and LAA measurements and the calculated (106) inferred metrics for the patient's LA and LAA using a classifier based on the trained machine learning includes: calculating the patient's stroke risk score based on the extracted (104) LA and LAA measurements and the calculated (106) inferred metrics for the patient's LA and LAA using a trained deep neural framework.

Images