Method, device and computer-readable medium for the automatic detection of a hemodynamically significant coronary stenosis

Description

[0001] The present invention relates to a computer-implemented method for automatically determining the presence of a hemodynamically significant coronary stenosis by predicting the FFR (fractional flow reserve) value associated with the stenosis thus detected, as well as a device capable of automatically determining the presence of a hemodynamically significant coronary stenosis by predicting the FFR (fractional flow reserve) value associated with the stenosis thus detected and a non-transitory computer-readable medium storing computer-readable program instructions for automatically determining the presence of a hemodynamically significant coronary stenosis by predicting the FFR value associated with the stenosis thus detected.

[0002] Coronary artery disease is the second leading cause of death in developed countries, after cancer. It affects more than fifteen million Americans. It can manifest itself suddenly; it is by far the leading cause of sudden death worldwide. There are approximately 50,000 cases of sudden death per year in France, most of them from myocardial infarction. It can affect young people, sometimes in their thirties. Its incidence increases with the aging of the population and the development of chronic diseases such as diabetes or high blood pressure.

[0003] Cardiovascular diseases (CVD) include a number of disorders affecting the heart and blood vessels such as: high blood pressure (hypertension); coronary heart disease (heart attack or infarction); cerebrovascular disease (stroke); peripheral arterial disease; heart failure; rheumatic heart disease; congenital heart disease; cardiomyopathy.

[0004] Coronary heart disease, also called coronary artery disease, coronary artery disease, or coronary insufficiency, is an obstructive disease of the coronary arteries, which supply blood to the heart.

[0005] When it progresses to stenosis (narrowing to occlusion), a coronary lesion leads to coronary artery disease, or coronary artery disease, or coronary insufficiency.

[0006] Coronary insufficiency generally results in myocardial ischemia, i.e. insufficient blood supply (ischemia) to the heart muscle (myocardium), due in particular to vascular obstruction.

[0007] Many additional examinations can be used to explore myocardial ischemia, the main ones being the electrocardiogram, stress test, MRI, myocardial scintigraphy, and coronary angiography, and more recently coronary CT angiography.

[0008] In practice, the classification of coronary stenoses according to their severity is most often carried out visually by coronary CT angiography: it depends on a segmentation based on the extraction of the central line of the vessel. It is dependent on the experience of the reader. A stenosis is usually considered significant for a reduction in diameter of at least 50%, by visual estimation. This visual assessment remains imprecise with substantial inter-observer variability.

[0009] Accuracy in severity classification is closely related to image quality. Image quality depends on heart rate, possible stair-step artifacts between beats, the quality of contrast injection, the level of noise in the image, and the possible presence of severe calcifications. The latest cardiac CT technology allows for the best image quality by reducing most of the artifacts mentioned.

[0010] Extensive reading experience is required for coronary CT angiography (Kerl et al., "64-Slice Multidetector-row Computed Tomography in the Diagnosis of Coronary Artery Disease: Interobserver Agreement Among Radiologists With Varied Levels of Experience on a Per-patient and Per-segment Basis." J Thorac Imaging. Jan 2012;27(1):29-35.

[0011] These situations can lead to the wrong conclusion that coronary stenosis is present, sometimes leading to the unnecessary performance of conventional, invasive and expensive angiography.

[0012] Coronary CT angiography (CCTA) is a newer, highly sensitive method for the non-invasive detection of patients with suspected coronary artery disease, with a very high negative predictive value (usually greater than 95%). Being the most sensitive method, it tends to be used in coronary disease screening as a first-line examination.

[0013] On the other hand, coronary CT angiography has a lower specificity (around 50-70%) due to frequent false-positive cases. Its positive predictive value of CT is therefore lower. False-positive cases are observed in particular in cases of coronary calcifications and / or in cases of motion artifacts during image acquisition. The latter can increase or even create (non-existent) stenoses on the image. Thus, reading expertise is necessary to minimize the number of false positives. Good expertise is acquired over several years (at least 5 years) for radiologists or cardiologists working in a center specialized in cardiac imaging.

[0014] The SCOT-HEART Investigators paper, "Coronary CT Angiography and 5-Year Risk of Myocardial Infarction." N Engl J Med. 6 Sept 2018, describes, among other things, that the use of coronary CT angiography can reduce the rate of heart attack and mortality compared to a standard assessment by a conventional stress test.

[0015] However, the growing use of coronary CT angiography is in practice hampered by the level of expertise required for reliable interpretation in current practice.

[0016] The advent of Artificial Intelligence (AI) makes it possible to consider transferring certain elements of medical expertise into algorithmic form. Machine learning tools, particularly neural networks (NNs), make it possible to reproduce expertise, which is widely used in the field of image recognition. This is why multiple projects are developing in the field of medical imaging.

[0017] Thus, the present invention consists in particular in adapting expertise in the reading of coronary CT angiography using AI techniques.

[0018] The publication Zreik M et al., "A Recurrent CNN for Automatic Detection and Classification of Coronary Artery Plaque and Stenosis in Coronary CT Angiography." IEEE Trans Med Imaging. 2018 discloses a first method for automatic detection of stenosis by machine learning. However, this method only uses stretched MPR (for multiplanar reconstruction) images, and analyzes the arteries by volume fragments, with a classification into three grades (normal, less than 50% and greater than 50%) without automated image quality assessment, and does not associate it with functional assessment.

[0019] Application US20150112182 discloses the prediction of FFR by the use of Artificial Intelligence (by machine learning) from any type of medical image, in general. In this document, the application of CT imaging is mentioned as an example, as is ultrasound or MRI. However, this document does not mention the type of CT image chosen for learning, the grouped analysis of multiple images of the same lesion under several incidences, the number of images for analysis, and / or the influence of image quality.

[0020] According to Tonino et al., "Fractional Flow Reserve versus Angiography for Guiding Percutaneous Coronary Intervention." N Engl J Med. 2009 Jan 15; FFR measurement during coronary angiography is a gold standard for determining whether coronary stenosis is functionally significant, with a threshold of 0.8. Indeed, it has been shown in patients with stable angina who had an FFR of less than or equal to 0.8 that coronary angioplasty significantly reduced the subsequent need for urgent revascularization. Otherwise, angioplasty did not provide a prognostic benefit. In another study, the frequency of major cardiac events after angioplasty was reduced by 28% if the intervention was decided on the basis of the FFR value (less than or equal to 0.8) compared to the visual interpretation of coronary angiography images alone.

[0021] Coronary computed tomography angiography (CCTA) is a sensitive method for detecting coronary stenoses, allowing in practice to rule out coronary stenosis when the examination is normal. On the other hand, coronary computed tomography angiography tends to overestimate the degree of coronary stenosis, responsible for a lower specificity of this tool. The difference between the visual degree of a stenosis and its hemodynamic effect (and therefore its functional character) is known and confirmed in recent studies comparing different estimates of stenoses with their FFR value: in the FAME trial, the FFR was greater than 0.8 for 63% of stenoses estimated to be intermediate (50-70%), and for 20% of cases for stenoses estimated to be visually severe (between 70 and 99%). ( Tonino et al., “Angiographic Versus Functional Severity of Coronary Artery Stenoses in the FAME Study.” J Am Coll Cardiol. June 2010.

[0022] Thus, a non-invasive method that would provide both reliable anatomical and functional information and reduce the need for invasive conventional angiography with possible invasive FFR measurement would be desirable.

[0023] The present invention relates to a prediction of FFR on Artificial Intelligence (AI) analysis applied to images of coronary stenosis observed in CT under multiple incidences, associated or not with a combination of at least two anatomical criteria.

[0024] Other approaches use machine learning techniques for FFR prediction using a large database of coronary anatomies generated from a model. This type of approach is notably disclosed in the paper by Tesche et al., "Coronary CT angiography derived morphological and functional quantitative plaque markers correlated with invasive fractional flow reserve for detecting hemodynamically significant stenosis." J Cardiovasc Comput Tomogr. May 2016. However, in this paper the coronary images used are not from real patients. Their prediction is based on the principle of fluid hemodynamics.

[0025] Furthermore, the paper Nakanishi et al., "Automated estimation of image quality for coronary computed tomographic angiography using machine learning." Eur Radiol. Sept 2018, describes the use of deep learning for the automatic assessment of image quality in CCTA (coronary CT angiography). However, in this paper, poor-quality examinations were few due to artificial preselection that does not correspond to daily practice. Furthermore, in this paper, only axial, coronal, and sagittal images were analyzed, but not MRI images. However, MRI images are the basis of radiological interpretation in current practice.

[0026] The paper Lossau et al., "Motion artifact recognition and quantification in coronary CT angiography using convolutional neural networks." Med Image Anal. Feb 2019, also describes the use of deep learning for automatic image quality assessment in CCTA (coronary CT angiography). This paper discloses the feasibility of deep learning to quantify cardiac motion artifacts. However, in this paper, the overall image quality is not analyzed (including noise, low contrast, or significant calcifications). It is the overall image quality that provides a diagnostic confidence index taking into account all the parameters interfering with the image.

[0027] The present invention relates to a computer-implemented method for automatically determining the presence of a hemodynamically significant coronary stenosis by predicting a range of the FFR value associated with the stenosis thus detected, as well as a device capable of automatically determining the presence of a hemodynamically significant coronary stenosis by predicting a range of the FFR value associated with the stenosis thus detected, and a non-transitory computer-readable medium storing computer-readable program instructions for automatically determining the presence of a hemodynamically significant coronary stenosis by predicting a range of the FFR value associated with the stenosis thus detected.

[0028] Published studies in CCTA (coronary CT angiography) rely on visual estimates of stenoses, at the 50% diameter threshold. The relevance of this detection is highly dependent on the observer and their level of reading experience.

[0029] To date, there is no reliable tool for automatic detection of coronary lesions at the 50% threshold available in current practice, due to the multiple factors interfering with interpretation. These multiple factors include the following: contrast, noise, cardiac or respiratory movement, anatomical variation, calcifications, which make interpretation difficult.

[0030] Artificial Intelligence techniques use statistical models capable of reproducing expertise, which takes a long time to acquire. By training a neural network on thousands of images labeled by an expert, it is possible to approach the level of this expert without modeling. a priori. It is therefore possible to offer automatic detection of coronary stenosis with performance close to that of an expert.

[0031] This is what the Applicant was able to validate. The ability to detect stenoses thus exceeds 90% in the validation data.

[0032] It is from this threshold of approximately 50% in diameter that a lesion risks limiting coronary flow. The stenosis is said to be significant in this case. We then speak of hemodynamically significant coronary stenosis.

[0033] Beyond detecting threshold stenosis, it is very important to know whether the stenosis is hemodynamic or not. Indeed, only stenoses that cause a pressure drop downstream of the stenosis should be treated mechanically by stenting or possibly by coronary bypass.

[0034] This hemodynamic effect can be measured by measuring the pressure drop at maximum hyperemia. This is the basis for calculating FFR.

[0035] Studies have shown that there is a benefit to mechanically treating patients with coronary stenosis only if the FFR is less than or equal to 0.8, reflecting a pressure drop downstream of the stenosis.

[0036] This is why it is important to be able to predict an FFR level of less than or equal to 0.8 in the face of an image of coronary stenosis. Such a prediction makes it possible to avoid other more or less invasive and costly examinations whose performance is not always correlated with the results of invasive FFR.

[0037] If the FFR starts to decrease for stenoses from 40-50%, it is recognized that the degree of stenosis on the image does not allow to correctly and reliably predict the value of the FFR.

[0038] Between 50 and 70% of stenoses, two-thirds of patients have FFRs greater than 0.8, so their lesions are not hemodynamically significant. Between 70 and 99% of stenoses, however, 80% of patients have an FFR less than or equal to 0.8, so the stenosis is hemodynamically significant.

[0039] Thus, in the case of intermediate stenosis, i.e. between 50 and 80%, it is difficult to know whether or not to treat the patient with a mechanical system (stent), or by bypass. Indeed, to date, it is not possible to determine the hemodynamic nature or not of a stenosis on anatomical imaging alone.

[0040] Other anatomical criteria have been proposed in the literature but they do not seem to be sufficiently reliable to predict or not a hemodynamic effect of a stenosis.

[0041] These anatomical criteria include the degree of stenosis in surface or diameter, the minimum diameter, the minimum luminal surface or the length of the stenosis. However, these anatomical criteria do not appear to have been studied in association but only in isolation.

[0042] The Applicant was able to demonstrate that certain combinations of anatomical criteria were very relevant for determining the FFR value above or below the threshold of 0.8. This determination makes it possible, in particular, to opt for or not for treatment by stent or bypass.

[0043] Furthermore, directly learning a neural network from images associated with FFR values also allows such prediction on new images. Anatomical and artificial intelligence predictions appear complementary, and their combination seems very effective.

[0044] Stress imaging (ultrasound, MRI or CT) is used to highlight the hemodynamic nature of a stenosis. Other techniques based on coronary CT angiography alone, without using stress imaging, have also been proposed: use of the anatomical characteristics of the stenosis, without obtaining conclusive results, because the criteria were studied separately and not in combination as in the present invention; the attenuation gradient beyond the stenosis (TAG for transluminal attenuation gradient), which is not used in practice; an estimation of the FFR from a fluid dynamics model (Computational Fluid Dynamics or CFD). This last model is known as FFR-CT (for Fractional flow reserve-computed tomography). Simulation techniques require significant computing resources, the results of FFR-CT require 24 hours of processing. In addition, the results depend on the image quality, up to 20% of cases have been excluded from FFR-CT studies due to insufficient image quality.

[0045] Other methods exist but also have drawbacks. These include, for example, methods using pharmacological stress (Adenosine, Dypiridamole, or Dobutamine), such as stress MRI, which is expensive, time-consuming, and not widely available. Furthermore, gadolinium injection could expose patients to long-term risks, according to recent publications.

[0046] Stress scanning is the most recent application, which has not yet been evaluated on a large scale. It involves additional costs, iodine injection, and irradiation, which can be significant due to the repetitive acquisitions required to obtain a perfusion curve.

[0047] Nuclear medicine imaging also involves a radiation dose to the patient related to the injected radiotracer.

[0048] Stress ultrasound requires a trained operator, and its sensitivity appears to be low.

[0049] Considering the above, a problem that the present invention proposes to solve consists in particular in automatically detecting a potentially significant stenosis with good reliability and then predicting the FFR directly on anatomical images, when the stenosis is visually potentially significant, without the need for complex modeling or the addition of other imaging examinations. This has the particular advantage of being able to achieve considerable savings at the level of health systems throughout the world by greatly simplifying the patient's care pathway.

[0050] The solution to this problem has as its first object a computer-implemented method for determining the presence of coronary stenosis for a patient, comprising: a step of receiving at least one curvilinear or stretched multiplanar medical image of computed tomography (X-ray scanner) of said patient including the coronary stenosis; a step of detecting said coronary stenosis on said image or on a part of said image by the use of a first trained deep neural network; characterized in that it further comprises: a step of predicting a coronary flow reserve (FFR) value interval by manual, semi-automated and / or automated measurement of at least two morphological criteria chosen from: the minimum diameter of the stenosis in mm; the minimum surface area of the stenosis in mm 2<; the maximum degree of coronary stenosis expressed as a percentage (%) of diameter; the maximum degree of coronary stenosis expressed as a percentage (%) of surface area; the length of the stenosis in mm; and / or the myocardial mass or the percentage (%) of myocardial mass downstream of the coronary stenosis and / or a step of predicting a coronary flow reserve (FFR) value interval by using a second trained deep neural network, applied directly to the detected images or parts of images.

[0051] Its second object is a device capable of determining the presence of coronary stenosis for a patient, comprising: means for receiving at least one curvilinear multiplanar medical image of computed tomography (X-scan), of said patient including coronary stenosis; means for detecting said coronary stenosis on said image or on a part of said image, by the use of a first deep neural network; characterized in that it further comprises means for predicting a coronary reserve flow value interval by using a second trained deep neural network, applied directly to the detected images or parts of images.

[0052] Finally, the invention provides a non-transitory computer-readable medium storing computer-readable program instructions for determining the presence of coronary stenosis for a patient, comprising executing by a processor computer-readable program instructions having the effect of performing the following operations: receiving at least one curvilinear multiplanar medical computed tomography (X-ray) image of said patient including coronary stenosis; detecting said coronary stenosis on said image or on a portion of said image by using a first trained deep neural network; characterized in that it further generates the performance by said processor of an operation of prediction of a coronary reserve flow value interval by use of a second trained deep neural network, applied directly to the detected images or parts of images.

[0053] The Applicant has notably been able to develop a method which has qualities enabling the automatic detection of potentially significant coronary stenosis from a hemodynamic point of view, on the basis of high-level expertise. In addition, the method has the advantage of being able to predict the hemodynamic nature of the stenosis by predicting an FFR threshold with a high probability. This system therefore makes it possible, in a single coronary CT angiography examination, to provide reliable results which can advantageously subsequently be used to adapt therapeutic management.

[0054] The benefit for the patient compared to existing technologies is therefore very significant, since it provides diagnostic confirmation but above all prognostic value as well as appropriate therapeutic management, without resorting to other tests. Such a wealth of information makes it possible, in a second step, to simplify the cardiological diagnostic pathway in an obvious way. Furthermore, by reducing the need for other tests, the potential savings in terms of health expenditure are considerable.

[0055] In this description, unless otherwise specified, it is understood that, when an interval is given, it includes the upper and lower bounds of said interval.

[0056] The invention and the advantages resulting therefrom will be better understood upon reading the description and the non-limiting embodiments which follow, illustrated with reference to the appended drawings in which: There figure 1 represents a curvilinear MRI image of a coronary artery showing stenosis. figure 2 represents a stretched MRI image of a coronary artery showing stenosis. The figure 3 is a functional diagram illustrating the different possible stages of a method according to the invention. The figure 4 represents some of the anatomical criteria for predicting whether a stenosis is hemodynamic or not. The Figure 5 illustrates the automatic detection of a stenosis with an FFR less than or equal to 0.8 (FFR+) on a curvilinear MRI image by a second neural network. The figure 6 illustrates the performance of lesion detection with FFR less than or equal to 0.8 (FFR+) by the second neural network, for 1 given image. The figure 7illustrates the analysis of multiple images of the same lesion observed under different incidences, allowing a more robust and more efficient classification of the FFR at the threshold of 0.8, compared to the analysis of an isolated image. The techniques of majority vote or the highest average of the classifications are efficient. 9 images of the same artery, seen under different incidences (shifted by at least 20°), are analyzed successively. The 9 images are classified by the neural network as “FFR+”. The final overall classification is therefore “FFR+” due to the majority classification (9 / 9) and the highest average (0.643) of the “FFR+” classification. figure 8illustrates the analysis of multiple images of the same lesion observed under different incidences, allowing a more robust and more efficient classification according to the CAD-RADS classification, compared to the analysis of an isolated image. The techniques of majority vote or the highest average of the classifications are efficient. The figure 8 shows 9 images of the same artery, seen from different incidences (shifted by at least 20°), which are analyzed successively. 7 images are classified by the neural network as “normal” (CAD-RADS 0) 2 as “obstructive” (CAD-RADS 3 or 4). The overall final classification is therefore “normal” (CAD-RADS 0) due to the majority classification (7 / 9) and the highest average (0.605) on the 9 images. figure 9illustrates the analysis of multiple image fragments of the same lesion observed under different incidences, allowing a robust and efficient classification of the FFR at a threshold value of 0.8. The clear zone corresponds to the analyzed image fragment, limited to the coronary lesion, advantageously allowing a higher performance of analysis by the neural network. Here the highest average is 0.503 for FFR +. The final classification of the lesion is therefore “FFR +”.

[0057] The first subject of the invention is a computer-implemented method for determining the presence of coronary stenosis for a patient.

[0058] The first step of said method is a step of receiving at least one curvilinear or stretched multiplanar medical image of computed tomography (X-ray scanner) of said patient including the coronary stenosis. A multiplanar image is an image reconstructed from the center line of a tubular anatomical structure such as a vessel, for example a coronary artery. The major axis of the image plane is then aligned with an anatomical structure by following this center line. This makes it possible to include the entire anatomical structure (here a coronary artery) in a single image. An MPR image can follow the curvilinear path of the vessel, the adjacent structures are then distorted. The axis of the vessel can be stretched by projection in a fixed direction. The visualization can be done on a 360° rotation axis in both modes, curvilinear or stretched.

[0059] The second step of the method is a step of detecting said coronary stenosis on said image or on a part of said image by using a first deep neural network.

[0060] Preferably, the images or parts of images come from a coronary CT angiography (or CCTA for Coronary Computed Tomography Angiography).

[0061] The first neural network is trained to read curvilinear MPR (multiplanar reconstruction) images, single images or for greater precision multiple MPR images of the same stenosis from several angles, ideally nine angles of 20° or more separation, to allow views covering a minimum field of 180°. A database of at least 5000, preferably 10000 coronary artery images was used, with stenoses classified as potentially hemodynamically significant based on information provided by an expert with more than 20 years of experience.

[0062] Curvilinear or stretched multiplanar images (MPR or multiplanar reconstructions or MPR) are generally used for the analysis of coronary arteries. They are obtained from the centerline of a coronary artery. This centerline is extracted by common software on radiology workstations, but it is sometimes necessary to manually correct the centerline so that this line always remains in the center of the circulating lumen. Each coronary artery is generally analyzed with multiple MPRs by multiplying the incidences over 180 or 360°. The method was developed and validated with nine images with a minimum of 20° of separation, thus covering at least 180°. This allows, in particular, to detect coronary plaques with greater sensitivity and to quantify coronary stenoses with greater accuracy.

[0063] A known method for automatic detection uses only stretched MRI images, analyzes arteries by volume fragments, with a classification into 3 grades (normal, less than 50% and greater than 50%), without automated image quality assessment.

[0064] For multiplanar reconstructions (MPR), the raw data are first reconstructed along a plane perpendicular to the z-axis, which is itself parallel to the patient's long axis. The resulting image is therefore a transverse or axial section of a three-dimensional organ. From this, one or more reconstruction planes can be selected. The resulting planes can be arbitrary, for example, frontal, coronal and / or sagittal, but they can be oriented along anatomical or lesion axes. A curvilinear landmark following the centerline of a vessel such as a coronary artery can be used to spread this structure along its entire length on a planar reconstruction.

[0065] A specific algorithm is used to classify a potentially hemodynamically significant lesion from multiple images. The algorithm calculates the most frequent classification of images of the same coronary artery according to different incidences, as well as the average of the probabilities of each classification. In the event of classification discordance, the final classification retains the most severe lesion (FFR+) in order to minimize the risk of wrongly not treating a patient (risk of false negative).

[0066] In the average calculations, the probability scores of each image lower than 0.2 are excluded because they are considered to be poorly discriminating by the neural network.

[0067] As is clear in particular from the Figure 5, a probability displayed by the neural network of 0.99 indicates that the lesion is very probably with an FFR less than or equal to 0.8, classification confirmed by the real data of FFR measured invasively in this patient (FFR=0.56). Furthermore, as is evident in particular from the figure 6 , at the optimal decision threshold of 0.35, the overall classification accuracy is 89.6%. The neural network thus appears very efficient in determining the FFR below the threshold of 0.8.

[0068] The method according to the invention further comprises: a step of predicting a coronary flow reserve (FFR) value interval by manual, semi-automated and / or automated measurement of at least two morphological criteria chosen from: the minimum diameter of the stenosis in mm; the minimum surface area of the stenosis in mm 2<; the maximum degree of coronary stenosis expressed as a percentage (%) of diameter; the maximum degree of coronary stenosis expressed as a percentage (%) of surface area; the length of the stenosis in mm; and / or the myocardial mass or the percentage (%) of myocardial mass downstream of the coronary stenosis; and / or a step of predicting a coronary flow reserve value interval by using a second trained deep neural network, applied directly to the detected images or parts of images.

[0069] According to a first embodiment of the invention, the method comprises only an additional step of predicting a coronary reserve flow value interval by using a second trained deep neural network, applied directly to the detected images or parts of images.

[0070] According to a second embodiment of the invention, the method comprises only an additional step of predicting a coronary reserve flow value interval by manual, semi-automated and / or automated measurement of at least two morphological criteria chosen from: the minimum diameter of the stenosis in mm; the minimum area of the stenosis in mm 2<; the maximum degree of coronary stenosis expressed as a percentage (%) of diameter; the maximum degree of coronary stenosis expressed as a percentage (%) of area; the length of the stenosis in mm; and / or the myocardial mass or the percentage (%) of myocardial mass downstream of the coronary stenosis.

[0071] According to a third embodiment of the invention, the method comprises both an additional step of predicting a coronary reserve flow value interval by using a second trained deep neural network, applied directly to the detected images or parts of images, and also an additional step of predicting a coronary reserve flow value interval by manual, semi-automated and / or automated measurement of at least two morphological criteria chosen from: the minimum diameter of the stenosis in mm; the minimum area of the stenosis in mm 2<; the maximum degree of coronary stenosis expressed as a percentage (%) of diameter; the maximum degree of coronary stenosis expressed as a percentage (%) of area; the length of the stenosis in mm; and / or the myocardial mass or the percentage (%) of myocardial mass downstream of the coronary stenosis.

[0072] Preferably, when the lesion is considered potentially hemodynamically significant, then anatomical criteria are extracted from the image, manually, semi-automatically or automatically: a: the minimum diameter of the stenosis in mm, b: the minimum area of the stenosis in mm 2< , c: the maximum degree of coronary stenosis expressed as a percentage (%) of diameter, d: the maximum degree of coronary stenosis expressed as a percentage (%) of area, e: the length of the stenosis in mm, f: the myocardial mass or the percentage (%) of myocardial mass downstream of the coronary stenosis.

[0073] Manual image extraction means manual measurement of the minimum vessel diameter and area (on an artery cross-section image) at the narrowest point of the stenosis; manual tracing or contouring of the diameter and area (on an artery cross-section image) at a segment of healthy artery closest to the stenosis upstream and downstream of the stenosis; manual measurement of the length of the stenosis; calculation of myocardial mass downstream of a stenosis by segmentation of the myocardium downstream of that stenosis; visual estimation of the percentage of myocardium downstream of a stenosis.

[0074] Semi-automatic image extraction refers to a measurement obtained by pre-creating centerlines from the user's pointing of a vessel. At each point of the vessel, the values of the minimum diameter of the vessel surface are displayed by an algorithm on the radiology workstation. The various parameters of interest are readable at the area of interest with the possibility of manual correction of the centerlines and contours. Specific dedicated algorithms can calculate the volume of vascularized myocardium downstream of a stenosis (depending on the dedicated radiology workstation).

[0075] Automatic image extraction refers to a measurement obtained automatically by automatically creating the lines and contours of the vessel when loading patient images. The measurements are then automatically generated by software. At each point of the vessel, the values of the minimum diameter of the vessel surface are displayed by an algorithm on the radiology workstation. The various parameters of interest are readable at the level of the area of interest with the possibility of manual correction of the central lines and contours. Specific dedicated algorithms can calculate the volume of vascularized myocardium downstream of a stenosis (depending on the dedicated radiology workstation).

[0076] Advantageously, the combination of at least two of these criteria and the neural network evaluation provides a prediction of the functional character by the FFR above or below the threshold of 0.8.

[0077] Preferably, the most relevant anatomical criteria for predicting an FFR value are: the minimum area of stenosis in mm 2< and the maximum degree of coronary stenosis expressed as a percentage (%) of area.

[0078] More preferably still, the most relevant anatomical criteria for predicting an FFR value are: the minimum area of stenosis in mm 2< , the maximum degree of coronary stenosis expressed as a percentage (%) of area, and the myocardial mass downstream of the stenosis.

[0079] Other criteria can also be used to predict an FFR value, these are the following criteria: the minimum diameter of stenosis in mm and the maximum degree of coronary stenosis expressed as a percentage (%) of diameter.

[0080] As illustrated in the figure 4 , the anatomical criteria for qualifying a stenosis are: Minimum diameter: D Minimum area: S Degree of stenosis in diameter: D / (D1-D2 / 2) Degree of stenosis in area: S / (S1-S2 / 2) Length of stenosis: L the myocardial mass or % of myocardial mass downstream of a stenosis

[0081] The Applicant was able to show, surprisingly, that, on a sample of 120 stenoses, at least one of these combinations was capable of separating 100% of lesions above or below the threshold of 0.83, a value very close to the clinically validated threshold of 0.8.

[0082] Advantageously, the step of predicting a coronary reserve flow value interval further comprises the use of a second trained deep neural network, applied directly to the detected images or parts of images.

[0083] The second neural network is trained to read curvilinear or stretched MPR images, single images or, for greater accuracy, multiple MPR images of the same stenosis in several incidences. The neural network was successfully trained on coronary images of patients for which the actual FFR value was measured invasively by intracoronary pressure sensor.

[0084] The analysis of multiple images of a coronary stenosis, seen from different incidences spaced 20° or more apart, is advantageous for a better classification by Artificial Intelligence of lesions according to the FFR threshold. Indeed, the appearance of a coronary lesion varies according to the viewing incidence: a lesion may be classified FFR+ according to one incidence, and FFR- according to another incidence. The FFR+ or FFR- classification of the same image will therefore be different according to each incidence. Compared to the analysis of a single image, the global analysis of FFR information on nine incidences increases the robustness and diagnostic accuracy (observed gain in accuracy of approximately 10%). For this, the principle of majority vote of each FFR classification is used, and / or the average of the classification scores of each image ( Fig. 7). To improve robustness, some images that have not reached a minimum FFR probability threshold for each category (e.g. a probability < 0.2) are excluded from the majority vote or the average.

[0085] Analysis by combining anatomical criteria is added to the analysis by Artificial Intelligence. If both analyses classify the lesion in the same way, this classification is proposed to the user.

[0086] In the event of a discrepancy between anatomical criteria and Artificial Intelligence criteria, the FFR+ classification is proposed to reduce the risk of False Negatives, because it is considered more serious not to diagnose a treatable lesion than to overestimate a lesion (this thus promotes the sensitivity of the diagnosis of hemodynamic stenoses).

[0087] So, for example: a / Nine MPR images of the same artery with stenosis identified by the first neural network are entered into the algorithm. It finds five FFR images classified as “+” and four FFR images classified as “-”. The average probability of FFR “+” is 0.6, that of FFR “-” is 0.5. In this case, the lesion is classified as FFR “+” (hemodynamically significant) because the FFR “+” classification is more frequent and the average probability of FFR “+” is higher. The lesion will be classified as FFR “+” with a probability of 0.6. b / Seven other MPR images of a second artery with stenosis identified by the first neural network are entered into the algorithm. Six are FFR “-” with an average probability of 0.9, the last is FFR “+” with an average probability of 0.7. The classification is most frequently FFR “-” with the highest probability. The stenosis is then judged to be non-hemodynamically significant (with a high probability: 0.9).

[0088] In the average calculations, probability scores lower than 0.2 are excluded because they are considered to be poorly discriminating by the neural network.

[0089] The result of this second neural network, if it confirms the first prediction based on anatomical criteria, makes the prediction highly probable. In the event of a discrepancy, the result of the FFR+ prediction takes precedence in order to prioritize detection sensitivity by minimizing the number of false negatives. It is indeed considered more serious to underestimate a lesion than to overestimate it.

[0090] The method according to the invention can advantageously also comprise a step of determining a value according to the CAD-RADS (Coronary Artery Disease - Reporting and Data System value) classification of a coronary stenosis by using a third trained deep neural network, applied directly to the detected images or parts of images.

[0091] The said third neural network was successfully trained by supervised learning on arteries in which a CAD-RADS classification was applied by a recognized expert.

[0092] A specific algorithm allows the classification of a coronary artery according to CAD-RADS from multiple images. The algorithm takes the most frequent classification of images from different incidences (from 0 to 5), and compares it to the average of the probabilities of each classification. If the most frequent classification is also the one with the highest average probability, this is retained by the algorithm. In case of discordance, the most frequent classification is retained. The algorithm eliminates from the calculations the probability scores below a certain decision threshold, previously defined in order to optimize the diagnostic performance of the neural network. The analysis of multiple images of a coronary stenosis, seen from different incidences spaced 20° or more apart, is advantageous for a better classification of lesions according to their CAD-RADS classification.Indeed, coronary lesions (plaques or stenoses) are often not developed symmetrically. Consequently, the appearance of the lesion varies depending on the proposed incidence: a lesion may appear tight according to one incidence, less tight or even sometimes non-existent according to a third incidence. The classification of the same image will therefore be different depending on its incidence. Compared to the analysis of a single image, the global analysis of CAD-RADS information on nine incidences increases robustness and precision (observed gain of 10% precision). For this, the principle of majority vote of each CAD-RADS classification per image is used, and / or the average of the classifications of each image. To improve robustness, certain images that have not reached a minimum probability threshold for each CAD-RADS category are excluded from the vote or the average.

[0093] So, for example: a / Nine MRI images of the same artery are classified by the third neural network. Five MRI images are classified CAD-RADS 4, two MRI images are classified CAD-RADS 3, two MRI images are classified CAD-RADS 2. The average probability of CAD-RADS 4 is 0.8, that of CAD-RADS 3 is 0.5, that of CAD-RADS 2 is 0.3. In this case, the lesion is classified CAD-RADS 4 because this classification is more frequent and its average probability CAD-RADS 4 is higher. The lesion will be classified CAD-RADS 4 with a probability of 0.8 (high confidence). b / Seven other MRI images of the same artery are classified by the third neural network. Four MRI images are classified as CAD-RADS 2, three MRI images are classified as CAD-RADS 3. The average probability for CAD-RADS 2 is 0.6, for CAD-RADS 3 it is 0.7. In this case, the lesion is classified as CAD-RADS 2 because this classification is more frequent. The lesion will be classified as CAD-RADS 2 with a probability of 0.6 (with low confidence).

[0094] In the average calculations, probability scores lower than 0.2 for a category are excluded because they are considered too indiscriminate by the neural network.

[0095] The CAD-RADS classification allows for a rational and standardized classification of atheromatous coronary lesions. This classification into 6 degrees of severity (from 0 to 5) makes it possible to propose optimal therapeutic choices for the patient in light of the results of the coronary CT scan.

[0096] To date, no automatic classification based on CAD-RADS has been proposed. Such automatic detection, if reliable, is particularly useful in facilitating the daily interpretation work of radiologists or cardiologists.

[0097] Advantageously, the CAD-RADS classification strengthens the predicted value of FFR, because CAD-RADS classifications lower than 3 correspond to stenoses lower than 50%.

[0098] Therefore, a CAD-RADS score less than 3 is probably not associated with an FFR less than or equal to 0.8.

[0099] Conversely, a CAD-RADS 4 classification strongly increases the probability of an FFR value less than or equal to 0.8.

[0100] The method according to the invention thus advantageously comprises a CAD-RADS classification for each curvilinear MRI image of the same artery from different angles.

[0101] For a group of one to nine images, the average probability score of each CAD-RADS category is calculated along with the most frequent CAD-RADS category. If the category with the highest score is also the most frequent category, it is retained as the most probable classification ( Fig. 8). Otherwise, the most severe result of the two classifications is retained, in order to minimize the risk of underestimating a lesion. Overestimation is considered more dangerous than overestimation for the detection of coronary lesions. CAD-RADS 0: normal CAD-RADS 1: plaque < 25% CAD-RADS 2: plaque between 25 and 49% CAD-RADS 3: 50-69% (stenosis) CAD-RADS 4: 70-99% (stenosis) CAD-RADS 5: occlusion

[0102] Preferably, for a more robust classification, categories 1 and 2 and categories 3 and 4 can be grouped as follows, according to the wishes of the end user. CAD-RADS 0: normal CAD-RADS 1 or 2: non-obstructive coronary artery disease CAD-RADS 3 or 4: obstructive coronary artery disease CAD-RADS 5: occlusion

[0103] The method according to the invention advantageously further comprises at least one of the following steps, which can be carried out in any order: a step of automated determination of image quality providing a diagnostic confidence index by using a fourth trained neural network, applied directly to the detected images or parts of images; a step of determining an overall calcification score on a scale of 0 to 4 predicting the category of the Agatston calcium score, by using a fifth trained neural network, applied directly to the detected images or parts of images; and / or a step of determining high-risk plaque (HRP) of cardiac event, by using a sixth trained neural network, applied directly to the detected images or parts of images.

[0104] Excellent image quality is advantageous for obtaining a reliable, relevant and accurate diagnosis in CCTA. The presence of artifacts, related to cardiac motion, insufficient contrast, or noise in the image (noise is measured as the standard deviation of pixel values in a homogeneous region of an image), interfere with the diagnosis and quantification of coronary stenosis and make subsequent therapeutic decisions more difficult. Automatic assessment of image quality is useful for quality control, and for comparing images from one center to another. In the present invention, this automatic assessment is advantageously used to provide a diagnostic confidence index in the final interpretation.

[0105] The fourth neural network for the automated image quality determination step was successfully trained by supervised learning on MRI images of arteries whose image qualities were assessed by a recognized expert. The images were classified according to a score of 0 to 4 according to the subjective scale (detailed below).

[0106] The neural network provides an overall image quality score from one to nine images of the same artery. The average value of the image quality scores is calculated. This value defines a confidence index between 0 and 4.

[0107] A specific algorithm allows to classify the image quality from multiple images of the same artery. The algorithm takes the average of the classifications of the images of an artery according to different incidences (classified from 0 to 4).

[0108] So, for example: a / Nine MPR images of the same artery are classified by the fourth neural network. Five MPR images are classified QI 4, four MPR images are classified QI 3.

[0109] The retained quality will be (5*4+4*3) / 9 =3.6. This figure is for example considered as a confidence indicator for the final analysis.

[0110] The image quality classification is listed below: IQ=0- Not assessable IQ=1- Poor image quality. Presence of artifacts. Low diagnostic confidence IQ=2. Fair. Interpretation is possible but the confidence level is low IQ=3. Good IQ. Good diagnostic confidence IQ=4. Excellent IQ. High confidence

[0111] The fifth network for the step of determining a global calcification score was trained directly on MRI images of arteries for which the Agatston score was known by a prior CT examination without injection of contrast agent. After training, the degree of calcification is predicted semi-quantitatively on an injected CT angiogram, according to four categories, for each of the extracted arteries: 0: No calcification 1: Moderate calcification: predicted Agatston score between 1 and 99 2: Moderate calcification: predicted Agatston calcium score: between 100 and 400 3: Severe calcification: predicted Agatston calcium score: greater than 400

[0112] Coronary calcium detection is documented in the literature: in particular, the Agatston Score is recognized as an important and independent risk marker for predicting coronary events, along with known risk factors such as high cholesterol levels, diabetes or hypertension.

[0113] An algorithm retains the highest score on images of the same artery according to multiple incidences, then adds up the scores obtained for each artery to obtain an overall calcification score, making it possible to predict a risk: Sum = 0 predicted Agatston calcium score: zero Sum = 1 predicted Agatston calcium score: 1-100 Sum = 2 predicted Agatston calcium score: 100-200 Sum = 3 predicted Agatston calcium score: 200-400 Sum >=4 predicted Agatston calcium score: > 400

[0114] In the literature, the Agatston score makes it possible to estimate the risk of cardiovascular events at 10 years: 0: minimal risk, Less than 100: low risk, 100-400: intermediate risk, > 400: high risk.

[0115] According to the method of the invention, the calcium score is calculated beforehand on a scanner without contrast injection, because the high-density contrast interferes with the detection of calcium. Machine learning was used to estimate the score automatically on examinations with contrast, having previously provided the score obtained on a scanner without contrast of the same patient.

[0116] So, for example: For a given patient, five images are analyzed for each major artery (IVA for the left anterior descending artery, Cx for the circumflex artery and CD for the right coronary)

[0117] The calcium scores are as follows: IVA: 0,0,0,0.1 Cx: 1,1,1,2,1 CD: 0,0,0,0,0

[0118] So the score Ca = Max IVA + Max Cx + Max CD = 1 + 2 + 0 = 3

[0119] With a score of 3, Agatston's predicted calcium score will be between 200 and 400, corresponding to an intermediate risk.

[0120] Finally, the sixth network for the high-risk plaque determination step was successfully trained by supervised learning on MPR images or cross-sectional images of arteries perpendicular to the MPR images (cross sectional images) in which the possible presence of a vulnerable plaque was or was not detected by a recognized expert. After training, the presence of a vulnerable plaque is affirmed according to a probability threshold (between 0 and 1) defined as the optimal threshold for the performance of this neural network.

[0121] High-risk plaques (HRP) are characterized by the presence of the following elements: low-density plaque (LDP), positively remodeling plaque (PR) by increase in the vessel wall towards the outside, presence of a negative density zone within the plaque (lipid core).

[0122] The plaque most often has an asymmetrical character.

[0123] The neural network uses MRI images which can be supplemented with images of artery sections at the plaque level (such as CROSS SECTIONAL IMAGES) to improve the accuracy of detection.

[0124] To date, there is no deep learning-based system that automatically detects at-risk plaques from MRI images.

[0125] The presence of a vulnerable plaque is retained if the score is higher than the threshold thus defined on at least one of the MPR images or on an artery section image.

[0126] A specific algorithm is used to determine the presence of a vulnerable plaque. Due to its asymmetric nature, a plaque may not be visible on one or more MRI image incidences due to a different viewing angle.

[0127] For example: a / Five MRI images of the same artery are analyzed by the sixth neural network. The presence of a vulnerable plaque is noted V, if the probability threshold reaches or exceeds 0.5, its absence is noted 0.

[0128] The result 0,0,0,V,0 corresponds to the presence of a vulnerable plaque (because the presence was detected on at least one MRI image).

[0129] Preferably, as illustrated in figure 3, the method which is the subject of the invention comprises the following five steps which can be carried out in any order: a step of predicting a coronary reserve flow value interval using a second trained deep neural network, applied directly to the detected images or parts of images; a step of determining a value according to the CAD-RADS classification using a third trained deep neural network, applied directly to the detected images or parts of images; a step of automated determination of image quality providing a diagnostic confidence index from a fourth trained neural network, applied directly to the detected images or parts of images; a step of determining an overall calcification score on a scale of 0 to 4 predicting the Agatston calcium score, using a fifth trained neural network, applied directly to the detected images or parts of images;and a step of determining a plaque at high risk of a cardiac event, using a sixth trained neural network, applied directly to the detected images or parts of images.;

[0130] The invention also relates to a device capable of determining the presence of coronary stenosis for a patient, comprising: means for receiving at least one curvilinear multiplanar medical image of computed tomography (X-ray scanner), of said patient including coronary stenosis; means for detecting said coronary stenosis on said image or on a part of said image, by the use of a first deep neural network.

[0131] Said device further comprises means for predicting a coronary reserve flow value interval by using a second trained deep neural network applied directly to the detected images or parts of images.

[0132] The first and second neural networks are as described above.

[0133] Preferably, the device further comprises means for determining a value according to the CAD-RADS classification by using a third trained deep neural network, applied directly to the detected images or parts of images.

[0134] The third neural network is as described above.

[0135] Advantageously, the device according to the invention further comprises at least one of the following means: automated means for determining image quality providing a diagnostic confidence index by using a fourth trained neural network, applied directly to the detected images or parts of images; means for determining an overall calcification score on a scale of 0 to 4 predicting the Agatston calcium score, by using a fifth trained neural network, applied directly to the detected images or parts of images; and / or means for determining plaque at high risk of cardiac event, by using a sixth trained neural network, applied directly to the detected images or parts of images.

[0136] The fourth, fifth and sixth neural networks are as described above.

[0137] According to a preferred embodiment of the invention, the device comprises the following five means: means for predicting a coronary reserve flow value interval using a second trained deep neural network, applied directly to the detected images or parts of images; means for determining a value according to the CAD-RADS classification using a third trained deep neural network, applied directly to the detected images or parts of images; automated means for determining image quality providing a diagnostic confidence index using a fourth trained neural network, applied directly to the detected images or parts of images; means for determining an overall calcification score on a scale of 0 to 4 predicting the Agatston calcium score, using a fifth trained neural network, applied directly to the detected images or parts of images;and means for determining high-risk plaque for cardiac events, by using a sixth trained neural network, applied directly to the detected images or parts of images.;

[0138] Finally, the invention provides a non-transitory computer-readable medium storing computer-readable program instructions for determining the presence of coronary stenosis for a patient, comprising executing by a processor computer-readable program instructions having the effect of performing the following operations: receiving at least one curvilinear multiplanar computed tomography (X-ray) medical image of said patient including coronary stenosis; detecting said potentially hemodynamically significant coronary stenosis on said image or on a portion of said image by using a first trained deep neural network; characterized in that it further generates the performance by said processor of an operation of prediction of a coronary reserve flow value interval by use of a second trained deep neural network, applied directly to the detected images or parts of images.

[0139] Preferably, the support is capable of further generating the performance by said processor of an operation of determining a value according to the CAD-RADS classification by using a third trained deep neural network, applied directly to the detected images or parts of images.

[0140] More preferably, the support is also capable of causing said processor to carry out at least one of the following operations: automated determination of image quality providing a diagnostic confidence index by using a fourth trained neural network, applied directly to the detected images or parts of images; determination of an overall calcification score on a scale of 0 to 4 predicting the Agatston calcium score, by using a fifth trained neural network, applied directly to the detected images or parts of images; and / or determination of plaque at high risk of cardiac event, by using a sixth trained neural network, applied directly to the detected images or parts of images.

[0141] According to a preferred embodiment of the invention, the support causes said processor to carry out the following five operations: prediction of a coronary reserve flow value interval by using a second trained deep neural network, applied directly to the detected images or parts of images; determination of a value according to the CAD-RADS classification by using a third trained deep neural network, applied directly to the detected images or parts of images; automated determination of image quality providing a diagnostic confidence index by using a fourth trained neural network, applied directly to the detected images or parts of images; determination of an overall calcification score on a scale of 0 to 4 predicting the Agatston calcium score, by using a fifth trained neural network, applied directly to the detected images or parts of images;and determination of plaque at high risk of cardiac event, by using a sixth trained neural network, applied directly to the detected images or parts of images.;

[0142] The present invention will now be illustrated by means of the following examples: Example 1:

[0143] A large database of at least 5,000, preferably more than 10,000 MPR images from coronary CT angiography was used for supervised learning. All images were classified and labeled by an expert with more than 20 years of experience reading these images (cumulative experience of approximately 50,000 cases analyzed). The images in this database were classified in terms of image quality, degree of calcification, and degree of stenosis according to the CAD-RADS classification. The possible presence of at-risk plaque (vulnerable plaque) was specified. In addition, on more than 2,000, preferably 4,000 images of patients with stenosis and known FFR value, a binary classification at the FFR threshold of 0.8 was performed. A neural network could thus be trained (on 80% of the images) to predict on a new image (among the 20% of the remaining images, serving as a test base) whether the FFR value will be greater or less (> or <) than 0.8.

[0144] Different neural networks available in open access were tested: GOOGLENET ™< , RESNET ™< and INCEPTION ™< V3, VGG11 ™< , VGG13 ™< , VGG19 ™< in order to obtain the best classification rate. Various measurements were carried out on a test database, independent of the training database: calculations of test accuracy, sensitivity, specificity, positive predictive value, F1 score (harmonic mean between sensitivity and positive predictive value), area under the ROC curve (Receiving Operator Curve, with sensitivity on the abscissa and 1-specificity on the ordinate).

[0145] An area under the curve reaching 0.85 was obtained, reflecting good overall predictive performance, equal to or better than current predictive systems. Scores are expected to improve with progressively increasing training base.

[0146] MPR multiplanar images are provided by the radiology post-processing consoles, based on central lines. Curvilinear or stretched MPR images, one to nine images of the same artery. Images can be exported from workstations in a standard image format (e.g., JPEG or PNG), and subsequently uploaded to a dedicated website. They can also be directly uploaded to a website from the workstation in standard radiology DICOM format. The resulting evaluation result is then returned to the reader's usual environment. Neural networks can also be directly integrated into post-processing consoles via dockerization. Example 2: Creation of neural networks, methods and results

[0147] A large database of stretched or curvilinear MRI images was created from examinations performed on different scanners.

[0148] Patient data has been anonymized.

[0149] For each dataset, MRI images of the three main arteries were extracted, namely: IVA, the left circumflex, and the right coronary artery.

[0150] For each of the three arteries, nine images were selected with different viewing angles (20° minimum difference between 2 images), by rotation around the central line.

[0151] Each MRI image was classified by an expert with regard to image quality, degree of calcification, presence or absence of vulnerable plaque, and degree of stenosis using the CAD-RADS classification.

[0152] A reduced database of more than 1800, preferably 4500 images on more than 200, preferably 500 patients with a known FFR value was also trained.

[0153] The images were loaded onto a platform (DEEPOMATIC ™< STUDIO, Paris, France and CLEVERDOC, Lille, France) allowing image classification and testing of the different neural networks.

[0154] For each neural network, 80% of the images in the database were used for training, and 20% of the images were used for evaluation. The evaluation images were excluded from the training process.

[0155] For each task, different neural networks were used for training.

[0156] The networks associated with the best results (best sensitivity and best positive predictive value) were selected. Results and conclusion:

[0157] For the automatic detection of vulnerable plates the average F1 score reached 60%.

[0158] For the automatic calcium score the average F1 score reaches 75%.

[0159] For automatic image quality classification the F1 score reaches 75%.

[0160] For the classification of stenoses according to CAD-RADS the average F1 score reaches 87%.

[0161] For the prediction of FFR at the threshold of 0.8, the average F1 score reaches approximately 85% or even 87%.

[0162] The results obtained on this first basis therefore seem equal to or superior to those of other methods which are more complex to implement and more expensive.

Claims

1. A computer-implemented method for determining the presence of a coronary stenosis in a patient, comprising: - a step of receiving multiplanar curved or stretched medical computed tomography (CT) images including the coronary stenosis of said patient, said images including representations of the coronary stenosis viewed under different angles; - a step of detecting said coronary stenosis in said images or in a portion of said images by using a first trained deep neural network configured to read said images; - a step of predicting a value interval of fractional flow reserve (FFR) by using a second trained deep neural network applied directly to said detected images or detected image portions.

2. Method according to claim 1, characterized in that it further comprises a step of determining a value according to the CAD-RADS (Coronary Artery Disease - Reporting and Data System) classification of a coronary stenosis by using a third trained deep neural network applied directly to the detected images or image portions.

3. Method according to claim 1 or 2, characterized in that it further comprises at least one of the following steps, which may be carried out in any order: - a step of automated image quality determination providing a diagnostic confidence index by using a fourth trained neural network applied directly to the detected images or image portions; - a step of determining a global calcification score on a scale from 0 to 4 predicting the Agatston calcium score category by using a fifth trained neural network applied directly to the detected images or image portions; and / or - a step of determining high-risk plaque (HRP) of cardiac event by using a sixth trained neural network applied directly to the detected images or image portions.

4. Method according to claim 3, characterized in that it further comprises a step of predicting a value interval of fractional flow reserve (FFR) by manual, semi-automated and / or automated measurement of at least two morphological criteria selected from: • minimum diameter of the stenosis in mm; • minimum area of the stenosis in mm2; • maximum degree of coronary stenosis expressed as a diameter percentage (%); • maximum degree of coronary stenosis expressed as an area percentage (%); • length of the stenosis in mm; and / or • myocardial mass or percentage (%) of myocardial mass downstream of the coronary stenosis.

5. Method according to any one of the preceding claims, characterized in that the images or image portions are derived from a coronary CT angiography (CCTA).

6. A device suitable for determining the presence of a coronary stenosis in a patient, comprising: - means for receiving multiplanar curved medical computed tomography (CT) images including the coronary stenosis of said patient, said images including representations of the coronary stenosis viewed under different angles; - means for detecting said coronary stenosis in said images or in a portion of said images by using a first trained deep neural network configured to read said images; characterized in that it further comprises means for predicting a value interval of fractional flow reserve by using a second trained deep neural network applied directly to said detected images or detected image portions.

7. Device according to claim 6, characterized in that it further comprises means for determining a value according to the CAD-RADS classification by using a third trained deep neural network applied directly to the detected images or image portions.

8. Device according to claim 6 or 7, characterized in that it further comprises at least one of the following means: - means for automated image quality determination providing a diagnostic confidence index by using a fourth trained neural network applied directly to the detected images or image portions; - means for determining a global calcification score on a scale from 0 to 4 predicting the Agatston calcium score by using a fifth trained neural network applied directly to the detected images or image portions; and / or - means for determining high-risk plaque of cardiac event by using a sixth trained neural network applied directly to the detected images or image portions.

9. Device according to claim 8, characterized in that it comprises the following five means: - means for predicting a value interval of fractional flow reserve by using a second trained deep neural network applied directly to the detected images or image portions; - means for determining a value according to the CAD-RADS classification by using a third trained deep neural network applied directly to the detected images or image portions; - means for automated image quality determination providing a diagnostic confidence index by using a fourth trained neural network applied directly to the detected images or image portions; - means for determining a global calcification score on a scale from 0 to 4 predicting the Agatston calcium score by using a fifth trained neural network applied directly to the detected images or image portions; and - means for determining high-risk plaque of cardiac event by using a sixth trained neural network applied directly to the detected images or image portions.

10. A non-transitory computer-readable medium storing computer-readable program instructions for determining the presence of a coronary stenosis in a patient, comprising execution by a processor of computer-readable program instructions configured to carry out the following operations: - receiving multiplanar curved medical computed tomography (CT) images including the coronary stenosis of said patient, said images including representations of the coronary stenosis viewed under different angles; - detecting said potentially hemodynamically significant coronary stenosis in said images or in a portion of said images by using a first trained deep neural network configured to read said images; characterized in that it further causes the processor to perform a prediction operation of a value interval of fractional flow reserve by using a second trained deep neural network applied directly to said detected images or detected image portions.

11. Medium according to claim 10, characterized in that it further causes the processor to perform an operation of determining a value according to the CAD-RADS classification by using a third trained deep neural network applied directly to the detected images or image portions.

12. Medium according to claim 10 or 11, characterized in that it further causes the processor to perform at least one of the following operations: - automated image quality determination providing a diagnostic confidence index by using a fourth trained neural network applied directly to the detected images or image portions; - determination of a global calcification score on a scale from 0 to 4 predicting the Agatston calcium score by using a fifth trained neural network applied directly to the detected images or image portions; and / or - determination of high-risk plaque of cardiac event by using a sixth trained neural network applied directly to the detected images or image portions.

13. Medium according to claim 12, characterized in that it causes the processor to perform the following five operations: - prediction of a value interval of fractional flow reserve by using a second trained deep neural network applied directly to the detected images or image portions; - determination of a value according to the CAD-RADS classification by using a third trained deep neural network applied directly to the detected images or image portions; - automated image quality determination providing a diagnostic confidence index by using a fourth trained neural network applied directly to the detected images or image portions; - determination of a global calcification score on a scale from 0 to 4 predicting the Agatston calcium score by using a fifth trained neural network applied directly to the detected images or image portions; and - determination of high-risk plaque of cardiac event by using a sixth trained neural network applied directly to the detected images or image portions.