Device for obtaining a trained machine learning segmentation model for pulvinar segmentation and device for diagnosis of status epilepticus

Abstract

The present invention relates to a device (1) for obtaining a trained machine learning segmentation model (31) for pulvinar segmentation. The invention also relates to a device (2) and method for diagnosis of Status Epilepticus (SE).

Description

DEVICE FOR OBTAINING A TRAINED MACHINE LEARNING SEGMENTATION MODEL FOR PULVINAR SEGMENTATION AND DEVICE FOR DIAGNOSIS OF STATUS EPILEPTICUSFIELD OF INVENTION

[0001] The present invention relates to the technical field of image processing and more particularly to a device for obtaining a trained machine learning segmentation model for pulvinar segmentation. The invention also relates to a device for diagnosis of Status Epilepticus (SE).BACKGROUND OF INVENTION

[0002] Status Epilepticus (SE) is characterized by prolonged or repeated seizures without a full recovery between them. SE may be observed in many diseases such as epilepsy, encephalitis, meningoencephalitis, intracranial hemorrhage, ischemic stroke, Traumatic Brain Injuries (TBI), brain tumors, meningitis, neurodegenerative diseases or Creutzfeldt- Jakob disease.

[0003] SE may be confused with other conditions that may appear to be SE such as stroke, low blood sugar, movement disorders, meningitis, or delirium. Misdiagnosing SE can have significant risks and consequences on the patient. Indeed, misdiagnosis can lead to a delay in providing the appropriate treatment. The main consequence is the increased risks of permanent neurological damages due to the repeated exposition to prolonged seizures.

[0004] In the meta-analysis of Nakae Y et al. (Relationship between cortex and pulvinar abnormalities on diffusion-weighted imaging in status epilepticus. J Neurol. 5016 Jan; 263(1): 127-32) it has been observed that in 42.5 % of the cases, SE is described in correlation with a pulvinar sign. The pulvinar sign is an hypersignal (such as a T2 signal hyperintensity) that may be observed in the pulvinar region on brain images for patients experiencing SE crisis. In most cases, the pulvinar sign is observed unilaterally, on only one of the two pulvinars.

[0005] Considering the low score of 42.5% from this meta- analysis, the pulvinar sign would not appear to be a convincing criterion for distinguishing SE from other conditions with similar symptoms.

[0006] However, in clinical routine, the pulvinar sign is only described in a qualitative manner, meaning that when medical professionals or researchers refer to the pulvinar sign, they are using descriptive terms and visual characteristics to identify and assess it, rather than relying on precise numerical measurements or quantitative data.

[0007] The work from Szabo K et al. (Diffusion-weighted and perfusion MR1 demonstrates parenchymal changes in complex partial status epilepticus. Brain. 5005; 128(6): 1369- 1376). demonstrated that when using a quantitative approach, rather than qualitative, for instance by measuring the Apparent Diffusion Coefficient (ADC) in the pulvinar region, 90% of the patients with SE showed a decrease in quantitative in ADC compared to surrounding regions of the thalamus.

[0008] This result highlights that the qualitative approach seems to be prone to a significant error rate due to the medical professional’s interpretation. Moreover, it shows that the quantification of the pulvinar sign could indeed be a convincing criterion for distinguishing SE from other conditions with similar symptoms.

[0009] Therefore, there is a need for an improved and reliable way of quantifying the pulvinar sign in brain imaging.SUMMARY

[0010] This invention thus relates to a device for obtaining a trained machine learning (ML) segmentation model for pulvinar segmentation, said device comprising: at least one input configured (i.e. adapted) to receive a plurality of image pairs (e.g., from a plurality of subjects), each pair of images comprising: o at least one ground-truth image, said ground-truth image being a T1 mapping image, and o at least one FLAIR (Fluid Attenuated Inversion Recovery) image,wherein the at least one ground-truth image and the at least one FLAIR image are obtained from a same subject (e.g., of said plurality of subjects); at least one processor configured to: o for each received pair of images, co-register the at least one ground- truth image and the at least one FLAIR image in a training reference space (i.e., same referential), and segment the pulvinar in the at least one ground-truth image so as to obtain at least one pulvinar segmentation mask (i.e., in the training reference space); o generate a training dataset comprising multiple training samples associated to a plurality of subjects, each training sample comprising the at least one pulvinar segmentation mask and the at least one FLAIR image from the same subject (e.g., of said plurality of subjects), wherein the at least one pulvinar segmentation mask and the at least one FLAIR image are co-registered in said training reference space; o obtain said trained machine learning segmentation model by training, using said training dataset, a convolutional neural network so as to obtain as output, a prediction of a pulvinar segmentation mask based on an input FLAIR image of a patient; at least one output configured to provide the obtained trained machine learning segmentation model.

[0011] According to the invention, co-registering images and / or masks into a reference space such as the training reference space means that the images and / or masks are transformed (i.e., a transformation is applied to the images and / or masks) so that they share the same spatial orientation and coordinate system. This allows for accurate and standardized comparisons, analysis, and measurements, ensuring that features in both images (and / or masks) are precisely matched, and data from different sources can be used together effectively. The co-registration in a training reference space between the at least one FLAIR image and the at least one ground truth image is performed on images captured on a same subject (e.g., of said plurality of subjects), which is why the coregistration is preferably a rigid registration. Advantageously, the co-registration in a training reference space allows to obtain images that are aligned as best as possible tooptimize the learning of the machine learning segmentation model on the at least one FLAIR image from the at least one ground truth image.

[0012] Advantageously, the device for obtaining a trained machine learning segmentation model for pulvinar segmentation from the present invention provides a trained machine learning segmentation model configured to generate a segmentation mask of the pulvinar using as input only routine magnetic resonance imaging (MRI) images representative of brain anatomical structures that are easily available for all patients. It has to be noted that these routine images (e.g., FLAIR images or T2FLAIR images) typically offer lower spatial resolution and less detailed anatomical information compared to T1 mapping images, thus rending accurate pulvinar segmentation more difficult or even impossible. The specific construction of the training dataset and training process proposed herein advantageously allows to obtain a robust machine learning segmentation model for pulvinar segmentation, capable of effectively differentiating between the pulvinar structure and surrounding tissues, even using as a starting point a FLAIR image. Indeed, the machine learning segmentation model for pulvinar segmentation performs a valuable, non-trivial task, made possible by its training using T1 mapping images, with "transfer" of the ability to segment the pulvinar onto FLAIR images through learning from databases that contain both T1 mapping images (e.g. MP2RAGE) and FLAIR images.

[0013] According to other advantageous aspects of the invention, the device comprises one or more of the features described in the following embodiments, taken alone or in any possible combination.

[0014] According to one embodiment, the at least one ground- truth image and the at least one FLAIR image are further co-registered into a standard space.

[0015] According to the present invention, a "standard space" is a common or standardized coordinate system or image space that may serve as a common framework for comparing and analyzing images from different individuals or sources. One of the most widely used standard spaces in neuroimaging is the MNI (Montreal Neurological Institute) space. The MNI space is a standardized, three-dimensional coordinate systemand brain template that serves as a reference for various neuroimaging studies. It provides a common framework for comparing and analyzing brain images from different individuals or sources.

[0016] Advantageously, the co-registration to a standard space allows for comparison between images captured on different subjects. Indeed, by putting the at least one ground truth image and the at least one FLAIR image in globally the same position for all the different subjects, it is possible to improve segmentation performances of the trained ML model, because it helps the convolutional neural network by giving invariant spatial priors. In other words, regardless of how images are transformed during co-registration to the standard space (such as translation, rotation, or scaling), the spatial priors (e.g. prior knowledge or assumptions about the spatial characteristics or properties of objects or regions within an image) remain stable and consistent. This co-registration to the standard space is preferably an affine registration.

[0017] The steps of co-registration of the at least one FLAIR image with the at least one ground-truth image in a training reference space and the co-registration of the at least one FLAIR image with the at least one ground-truth image to the standard space may be performed sequentially or simultaneously.

[0018] According to one embodiment, the co-registration of the at least one ground-truth image and the at least one FLAIR image in the training reference space is performed using trilinear interpolation.

[0019] According to one embodiment, the convolutional neural network is a 3D U-net.

[0020] Advantageously, a 3D U-net is the standard architecture for segmentation. It is based on an encoder-decoder structure but has the particularity to use skip connections that enable to keep the input image resolution, which is why it is one of the best performing architectures for segmentation.

[0021] According to one embodiment, during the training, the at least one processor is further configured to calculate, during the training of the ML segmentation model, a lossfunction based on (i.e., comprising at least) a sum of a smooth dice loss and a Binary Cross Entropy with logits.

[0022] According to another aspect, the invention relates to a device for pulvinar segmentation using the trained machine learning segmentation model for pulvinar segmentation obtained from the device for obtaining a trained machine learning segmentation model for pulvinar segmentation (described above), said device for pulvinar segmentation comprising: at least one input configured to receive at least one FLAIR image of the brain from a patient, at least one processor configured to provide as an input to said trained machine learning segmentation model said at least one FLAIR image so as to obtain at least one pulvinar segmentation mask, at least one output adapted to provide said pulvinar segmentation mask.

[0023] The disclosure also relates to a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out a computer-implemented method for training and / or a method for segmenting using the trained machine learning segmentation model, described above.

[0024] The disclosure also relates to a computer program comprising software code adapted to perform a method for training and / or a method for segmenting using the trained machine learning segmentation model, described above compliant with any of the above execution modes when the program is executed by a processor.

[0025] According to another aspect, the invention relates to a device for diagnosis of Status Epilepticus (SE) in a patient using a trained machine learning segmentation model for pulvinar segmentation obtained from a device (such as described above), said device comprising: at least one input configured to receive from said patient at least one FLAIR image of the brain, at least one B0 image of the brain, and at least one B1000 image of the brain; at least one processor configured to:• feed the at least one FLAIR image to the trained machine learning segmentation model to obtain at least one pulvinar segmentation mask;• co-register the at least one FLAIR image, the at least one BO image and the at least one B 1000 image in a reference space;• apply the obtained pulvinar segmentation mask to the at least one BO image (i.e., so as to obtain the segmented pulvinar region of the at least one BO image) and to the at least one B1000 image (i.e., so as to obtain a the segmented pulvinar region of the at least one B 1000 image);• calculate an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image, and / or calculate a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image;• compare said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion; and at least one output configured to output the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

[0026] Advantageously, the diagnosis device described above allows to accurately identify Status Epilepticus in a patient based on a non-invasive measurement of a signal such as Apparent Diffusion Coefficient or voxel intensity performed on a B0 image or a B 1000 image, wherein the pulvinar region is accurately segmented thanks to the pulvinar segmentation mask obtained with the trained machine learning segmentation model of the present invention. The diagnosis device may be particularly useful in the case of comatose patients for which SE is very difficult to diagnose because of the absence of abnormal movements during comas. The diagnosis device may also be used to better understand the occurrence of SE in diseases such as encephalitis.

[0027] According to one embodiment, the co-registration of the at least one FLAIR image, the at least one B0 image and the at least one B 1000 image in a reference space isperformed using rigid registration. Here, the co-registration is performed between images captured on a same patient.

[0028] According to one embodiment, the predefined criterion is an ADC calculated on the segmented pulvinar region of at least one BO image obtained on a healthy control subject and / or the intensity predefined criterion as a distribution of voxels intensity calculated on the segmented pulvinar region of at least one B1000 image obtained on a healthy control. In other words, the predefined criterion may comprise an ADC predefined criterion being an ADC calculated on the segmented pulvinar region of at least one BO image obtained on a healthy control subject and / or an intensity predefined criterion being a distribution of voxels intensity calculated on the segmented pulvinar region of at least one B1000 image obtained on a healthy control subject.

[0029] According to one embodiment, the subject is diagnosed with SE when the patient ADC is lower than said ADC calculated on the healthy control (i.e., ADC predefined criterion) and / or when the patient distribution of voxels intensity is shifted towards higher intensity values than said distribution of voxels intensity calculated on the healthy control (i.e., intensity predefined criterion).

[0030] According to the invention, the healthy control may be the contralateral pulvinar of the same subject, wherein the pulvinar sign is not observed or the pulvinar of another subject that is not suffering from SE.

[0031] According to another aspect, the invention relates to a computer- implemented method of diagnosing Status Epilepticus (SE) in a patient using a trained machine learning segmentation model for pulvinar segmentation obtained thanks to the device for obtaining a trained machine learning segmentation model for pulvinar segmentation (described above), said method comprising:- receiving from said patient at least one FLAIR image of the brain, at least one B0 image of the brain, and at least one B1000 image of the brain; feeding the at least one FLAIR image to said trained segmentation model to obtain at least one pulvinar segmentation mask;co-registering the at least one FLAIR image, the at least one BO image and the at least one B 1000 image in a reference space; applying the at least one pulvinar segmentation mask to the at least one BO image and to the at least one Bl 000 image; calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one Bl 000 image; comparing said ADC and / or said distribution of voxels intensity to a predefined criterion; and outputting the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

[0032] According to another aspect, the invention relates to a device for assessing the severity of Status Epilepticus (SE) in a patient using a trained machine learning segmentation model for pulvinar segmentation obtained from a device (such as described above), said device comprising: at least one input configured to receive from said patient at least one FLAIR image of the brain, at least one B0 image of the brain, and at least one B1000 image of the brain; at least one processor configured to:• feed the at least one FLAIR image to the trained machine learning segmentation model to obtain at least one pulvinar segmentation mask;• co-register the at least one FLAIR image, the at least one B0 image and the at least one B 1000 image in a reference space;• apply the obtained pulvinar segmentation mask to the at least one B0 image (i.e., so as to obtain the segmented pulvinar region of the at least one B0 image) and to the at least one B1000 image (i.e., so as to obtain a the segmented pulvinar region of the at least one B 1000 image);• calculate an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image, and / or calculate a distributionof voxels intensity on the segmented pulvinar region of the at least one Bl 000 image;• compare said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion; and at least one output configured to output the severity of the SE based on said comparison of said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion.

[0033] According to another aspect, the invention further relates to a method of assessing the severity of Status Epilepticus (SE) in a patient, using the trained machine learning segmentation model for pulvinar segmentation obtained from the device described, said method comprising: receiving from said patient at least one FLAIR image of the brain, at least one B0 image of the brain, and at least one B 1000 image of the brain; feeding the at least one FLAIR image to said trained segmentation model to obtain at least one pulvinar segmentation mask; co-registering the at least one FLAIR image, the at least one B0 image and the at least one B1000 image in a reference space; applying the at least one pulvinar segmentation mask to the at least one B0 image and to the at least one Bl 000 image; calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image; comparing said ADC and / or said distribution of voxels intensity to a predefined criterion; outputting the severity of the SE based on said comparison of said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion.

[0034] According to one embodiment, the SE is considered severe if the ADC is decreased as compared to an ADC predefined criterion.

[0035] According to another aspect, the invention further relates to a method of treating or preventing Status Epilepticus (SE) in a patient in need thereof, using the trained machine learning segmentation model for pulvinar segmentation obtained from the device described, said method comprising: receiving from said patient at least one FLAIR image of the brain, at least one BO image of the brain, and at least one B 1000 image of the brain; feeding the at least one FLAIR image to said trained segmentation model to obtain at least one pulvinar segmentation mask; co-registering the at least one FLAIR image, the at least one B0 image and the at least one B1000 image in a reference space; applying the at least one pulvinar segmentation mask to the at least one B0 image and to the at least one Bl 000 image; calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image; comparing said ADC and / or said distribution of voxels intensity to a predefined criterion; outputting the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion; and treating or preventing SE in the diagnosed patient.

[0036] According to one embodiment, the step of treating or preventing SE in the diagnosed patient comprises the administration of a treatment to said patient. According to one embodiment, the treatment is to be administered at a therapeutically effective amount to said patient.

[0037] According to one embodiment, said treatment may for instance be an anticonvulsant compound, an anti-seizure compound, or an antiepileptic compound.

[0038] According to one embodiment, said treatment may for instance be benzodiazepines, levetiracetam, sodium valproate, fosphenytoin, phenobarbital or propofol.

[0039] According to one embodiment, said treatment is selected from the group comprising or consisting of levetiracetam (CAS number 102767-28-2), sodium valproate (CAS number 1069-66-5), fosphenytoin disodium (CAS number 92134-98-0), phenobarbital (CAS number 50-06-6), propofol (CAS number 2078-54-8), lacosamide (CAS number 175481-36-4), ketamine (CAS number 6740-88-1) and benzodiazepines such as, for example clonazepam (CAS number 1622-61-3), diazepam (CAS number 439- 14-5), lorazepam (CAS number 846-49-1), clobazepam (CAS number 22316-47-8) or midazolam (CAS number 59467-70-8).

[0040] According to one embodiment, said patient is affected with a disease or condition selected from the group comprising or consisting of epilepsy, encephalitis, meningoencephalitis, intracranial hemorrhage, ischemic stroke, Traumatic Brain Injury (TBI), brain tumors, neurodegenerative diseases, and Creutzfeldt Jacob disease.

[0041] In addition, the disclosure relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the computer-implemented method of diagnosing Status Epilepticus (SE) described above.

[0042] The disclosure also relates to a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the computer-implemented method of diagnosing Status Epilepticus (SE) described above.

[0043] The disclosure also relates to a computer program comprising software code adapted to perform a method for diagnosing Status Epilepticus (SE) compliant with any of the above execution modes when the program is executed by a processor.

[0044] The present disclosure further pertains to a non-transitory program storage device, readable by a computer, tangibly embodying a program of instructions executable by the computer to perform a method of diagnosing Status Epilepticus (SE) compliant with the present disclosure.

[0045] Such a non-transitory program storage device can be, without limitation, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor device, or anysuitable combination of the foregoing. It is to be appreciated that the following, while providing more specific examples, is merely an illustrative and not exhaustive listing as readily appreciated by one of ordinary skill in the art: a portable computer diskette, a hard disk, a ROM, an EPROM (Erasable Programmable ROM) or a Flash memory, a portable CD-ROM (Compact-Disc ROM).DEFINITIONS

[0046] In the present invention, the following terms have the following meanings:

[0047] The terms “adapted” and “configured” are used in the present disclosure as broadly encompassing initial configuration, later adaptation or complementation of the present device, or any combination thereof alike, whether effected through material or software means (including firmware).

[0048] The term “processor” should not be construed to be restricted to hardware capable of executing software, and refers in a general way to a processing device, which can for example include a computer, a microprocessor, an integrated circuit, or a programmable logic device (PLD). The processor may also encompass one or more Graphics Processing Units (GPU), whether exploited for computer graphics and image processing or other functions. Additionally, the instructions and / or data enabling to perform associated and / or resulting functionalities may be stored on any processor- readable medium such as, e.g., an integrated circuit, a hard disk, a CD (Compact Disc), an optical disc such as a DVD (Digital Versatile Disc), a RAM (Random- Access Memory) or a ROM (Read-Only Memory). Instructions may be notably stored in hardware, software, firmware or in any combination thereof.

[0049] “Machine learning (ML)” designates in a traditional way computer algorithms improving automatically through experience, on the ground of training data enabling to adjust parameters of computer models through gap reductions between expected outputs extracted from the training data and evaluated outputs computed by the computer models.

[0050] A “hyper-parameter” presently means a parameter used to carry out an upstream control of a model construction, such as a remembering-forgetting balance in sample selection or a width of a time window, by contrast with a parameter of a modelitself, which depends on specific situations. In ML applications, hyper-parameters are used to control the learning process.

[0051] “Datasets” are collections of data used to build an ML mathematical model, so as to make data-driven predictions or decisions. In “supervised learning” (i.e. inferring functions from known input-output examples in the form of labelled training data), three types of ML datasets (also designated as ML sets) are typically dedicated to three respective kinds of tasks: “training”, i.e. fitting the parameters, “validation”, i.e. tuning ML hyperparameters (which are parameters used to control the learning process), and “testing”, i.e. checking independently of a training dataset exploited for building a mathematical model that the latter model provides satisfying results.

[0052] A “neural network (NN)” designates a category of ML comprising nodes (called “neurons”), and connections between neurons modeled by “weights”. For each neuron, an output is given in function of an input or a set of inputs by an “activation function”. Neurons are generally organized into multiple “layers”, so that neurons of one layer connect only to neurons of the immediately preceding and immediately following layers.

[0053] A “segmentation mask”, also known as a label map or voxel-wise mask, is a visual representation or binary image that assigns a unique label or class to each voxel in an input image. It is commonly used in image segmentation tasks to define and identify regions or objects of interest within an image. Each voxel in a segmentation mask corresponds to a specific region or object in the original image and is assigned a label or class value. Typically, a voxel with a value of 0 represents the background or unassigned region, while other non-zero values represent different classes or objects. Segmentation masks provide voxel-level annotations that facilitate the understanding and analysis of images by explicitly delineating regions of interest.

[0054] The above ML definitions are compliant with their usual meaning, and can be completed with numerous associated features and properties, and definitions of related numerical objects, well known to a person skilled in the ML field. Additional terms will be defined, specified or commented wherever useful throughout the following description.

[0055] “Treating” or “treatment”, as used herein, refers to alleviating a specified condition (such as Status Epilepticus), eliminating or reducing the symptoms of a condition (such as Status Epilepticus), slowing or eliminating the progression of a condition (such as Status Epilepticus), delaying the initial occurrence of a condition (such as Status Epilepticus) in a patient, or delaying the reoccurrence of a condition (such as Status Epilepticus) in a previously afflicted patient.

[0056] “Preventing”, as used herein, refers to preventing the initial occurrence of a condition (such as Status Epilepticus) in a patient, or preventing the reoccurrence of a condition (such as Status Epilepticus) in a previously afflicted patient.

[0057] “Diagnosing” or “diagnosis”, as used herein, refers to assessing the development or progression of a condition (such as Status Epilepticus). As is known to a person skilled in the art, the assessment can be accurately performed for a statistically significant subject, although it is intended to be accurate for 100% of the subjects to be diagnosed. Statistical significance can be easily determined by a person skilled in the art using methods widely known in the art, e.g., confidence interval determination, -valuc determination, / -test, Mann- Whitney test, and the like. Preferred confidence intervals are 90% or higher, 95% 10 or higher, 97% or higher, 98% or higher, and 99%. A preferred -valuc is 0.1, 0.05, 0.01, 0.005 or 0.0001. Preferably, diagnosis results according to the present invention will be accurate for 60% or more, 70% or more, 80% or more, or 90% or more of a group of subjects.

[0058] “Prognosing” or “prognosis”, as used herein, refers to predicating the outcome for a subject with a condition (such as Status Epilepticus), after a particular treatment or intervention.

[0059] “Subject”, as used herein, refers to individuals already diagnosed with the condition (Status Epilepticus) that may be used to generate a training dataset for training the machine learning segmentation model for pulvinar segmentation from the invention.

[0060] “Patient”, as used herein, refers to an individual to be diagnosed or treated according to the methods of the present invention. Subjects include, but are not limitedto, mammals (e.g., murines, simians, equines, bovines, porcines, canines, felines, and the like), preferably to primates, and most preferably to humans.

[0061] “Therapeutically effective amount” : refers to the level or amount of treatment as described herein that is aimed at, without causing significant negative or adverse side effects to the target, (1) delaying or preventing the onset of a disease, disorder, or condition; (2) slowing down or stopping the progression, aggravation, or deterioration of one or more symptoms of the disease, disorder, or condition; (3) bringing about ameliorations of the symptoms of the disease, disorder, or condition; (4) reducing the severity or incidence of the disease, disorder, or condition; or (5) curing the disease, disorder, or condition. A therapeutically effective amount may be administered prior to the onset of the disease, disorder, or condition, for a prophylactic or preventive action. Alternatively or additionally, the therapeutically effective amount may be administered after initiation of the disease, disorder, or condition, for a therapeutic action.

[0062] “Benzodiazepines”: refers to a group of compounds that comprise a “benzodiazepine nucleus” also known as “benzodiazepine core” which is the fusion of a benzene ring and a diazepine ring. Benzodiazepines facilitate the binding of the inhibitory neurotransmitter GABA on GABA receptors in the central nervous system (CNS) and therefore have a depressant effect on the CNS. Benzodiazepines are used in medicine as anticonvulsant, or for the treatment of anxiety or insomnia.BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The present disclosure will be better understood, and other specific features and advantages will emerge upon reading the following description of particular and non-restrictive illustrative embodiments, the description making reference to the annexed drawings wherein:

[0064] Figure 1 is a block diagram representing schematically a particular mode of a device for obtaining a trained machine learning segmentation model for pulvinar segmentation compliant with the present disclosure;

[0065] Figure 2 is a flow chart showing successive steps executed with the device for obtaining a trained machine learning segmentation model for pulvinar segmentation of Figure 1;

[0066] Figure 3 is a block diagram representing schematically a particular mode of a device for diagnosis of Status Epilepticus (SE) in a patient using the trained machine learning segmentation model for pulvinar segmentation obtained from the device of Figure 1;

[0067] Figure 4 is a flow chart showing successive steps executed with the device for diagnosis of Status Epilepticus (SE) from Figure 3;

[0068] Figure 5 is a block diagram representing schematically a particular mode of a device for assessing the severity of Status Epilepticus (SE) in a patient using the trained machine learning segmentation model for pulvinar segmentation obtained from the device of Figure 1;

[0069] Figure 6 is a flow chart showing successive steps executed with the device for assessing the severity of Status Epilepticus (SE) from Figure 5;

[0070] Figure 7 is a representation of a U-Net architecture of the machine learning segmentation model according to the invention;

[0071] Figure 8 shows examples of segmentations obtained using the trained machine learning segmentation model for pulvinar segmentation according to one embodiment of the invention based on FLAIR images and manual pulvinar segmentation based on T1 mapping images, for four patients from the testing set, at various slices, in FLAIR space;

[0072] Figure 9 shows a boxplot of the dice scores (ordinate: dice) on testing set between manual segmentations on ground truth images and segmentation obtained using the trained machine learning segmentation model for pulvinar segmentation according to one embodiment of the invention;

[0073] Figure 10 shows a regression between pulvinar volume segmented in ground truth images (abscissa : volume_gth) and with the trained machine learning segmentationmodel for pulvinar segmentation according to one embodiment of the invention (ordinate: volume_pred) for the 14 testing subjects;

[0074] Figure 11 shows a box plot of the relative difference between right and left pulvinar with manual segmentation;

[0075] Figure 12 shows the correlation for test subjects between mean pulvinar’ s ADC extracted with manual segmentation (abscissa : ADC_gth_mean_left) and with the trained machine learning segmentation model for pulvinar segmentation according to one embodiment of the invention (ordinate : ADC_prediction_mean_left) for left pulvinar; and

[0076] Figure 13 shows the correlation for test subjects between mean pulvinar’ s ADC extracted with manual segmentation (abscissa : ADC_gth_mean_right) and with the trained machine learning segmentation model for pulvinar segmentation according to one embodiment of the invention (ordinate : ADC_prediction_mean_right) for right pulvinar.ILLUSTRATIVE EMBODIMENTS

[0077] The present description illustrates the principles of the present disclosure. It will thus be appreciated that those skilled in the art will be able to devise various arrangements that, although not explicitly described or shown herein, embody the principles of the disclosure and are included within its scope.

[0078] All examples and conditional language recited herein are intended for educational purposes to aid the reader in understanding the principles of the disclosure and the concepts contributed by the inventor to furthering the art, and are to be construed as being without limitation to such specifically recited examples and conditions.

[0079] Moreover, all statements herein reciting principles, aspects, and embodiments of the disclosure, as well as specific examples thereof, are intended to encompass both structural and functional equivalents thereof. Additionally, it is intended that such equivalents include both currently known equivalents as well as equivalents developed in the future, i.e., any elements developed that perform the same function, regardless of structure.

[0080] Thus, for example, it will be appreciated by those skilled in the art that the block diagrams presented herein may represent conceptual views of illustrative circuitry embodying the principles of the disclosure. Similarly, it will be appreciated that any flow charts, flow diagrams, and the like represent various processes which may be substantially represented in computer readable media and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.

[0081] The functions of the various elements shown in the figures may be provided through the use of dedicated hardware as well as hardware capable of executing software in association with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, a single shared processor, or a plurality of individual processors, some of which may be shared.

[0082] It should be understood that the elements shown in the figures may be implemented in various forms of hardware, software or combinations thereof. Preferably, these elements are implemented in a combination of hardware and software on one or more appropriately programmed general-purpose devices, which may include a processor, memory and input / output interfaces.

[0083] The present disclosure will be described in reference to a particular functional embodiment of a device 1 for obtaining a trained machine learning segmentation model 31, as illustrated on Figure 1.

[0084] The device 1 is adapted to produce a machine learning segmentation model 31 that is trained for segmentation of the pulvinar. More precisely, the trained machine learning segmentation model 31 obtained from the device 1 is configured to receive as an input at least one FLAIR image 32 and to provide as output at least one pulvinar segmentation mask.

[0085] In the context of the invention, a FLAIR image is captured using a magnetic resonance imaging (MRI) scan combined with an inversion recovery technique, which involves selectively nullifying the signal from cerebrospinal fluid (CSF) while preserving the signal from other tissues. In this technique, an initial Radiofrequency pulse is applied to invert the magnetization of tissues, followed by a subsequent Radiofrequency pulse toselectively nullify the signal from CSF. For example, a FLAIR image 32 may be a T2FLAIR image.

[0086] These FLAIR images 32 have a lower contrast that usually make pulvinar segmentation challenging. Indeed, according to literature, FLAIR images may have a tendency to mislabel areas of high signal intensity in subcortical structures, as the pulvinar for instance, which may lead to inaccuracies. However, FLAIR images are highly available in clinical routine. Considering these facts, the device 1 is configured to output a trained machine learning model 31 capable of outputting a robust pulvinar segmentation mask capable of effectively differentiating between the pulvinar structure and surrounding tissues based on such low-contrasted FLAIR images 32.

[0087] To that end, the device 1 is configured to generate a training dataset and train a machine learning segmentation model 50 using said training dataset in order to obtain the trained segmentation model 31.

[0088] The device 1 for training the machine learning segmentation model 50 is associated with a device for pulvinar segmentation using the trained ML model obtained from device 1 and / or a device 2, represented on Figure 3, for diagnosis of Status Epilepticus (SE) in a patient using the trained machine learning segmentation model for pulvinar segmentation obtained from the device 1, which will be subsequently described. The device 1 may also be associated with a device 3 (represented on Figure 5) for assessing the severity of Status Epilepticus (SE) in a patient using the trained machine learning segmentation model for pulvinar segmentation obtained from the device 1.

[0089] Though the presently described devices 1 , 2 and 3 are versatile and provided with several functions that can be carried out alternatively or in any cumulative way, other implementations within the scope of the present disclosure include devices having only parts of the present functionalities.

[0090] Each of the devices 1, 2 and 3 is advantageously an apparatus, or a physical part of an apparatus, designed, configured and / or adapted for performing the mentioned functions and produce the mentioned effects or results. In alternative implementations, any of the device 1, the device 2 and the device 3 is embodied as a set of apparatus or physical parts of apparatus, whether grouped in a same machine or in different, possiblyremote, machines. The device 1 and / or the device 2 and / or the device 3 may have functions distributed over a cloud infrastructure and be available to users as a cloud-based service, or have remote functions accessible through an API.

[0091] The device 1 for training the segmentation model 50, the device 2 for diagnosis of Status Epilepticus (SE) in a patient and the device 3 for assessing the severity of Status Epilepticus (SE) in a patient may be integrated in a same apparatus or set of apparatus, and intended to same users. In other implementations, the structure of the device 2 (or the device 3) may be completely independent of the structure of the device 1, and may be provided for other users. For example, the device 2 (or the device 3) may have a trained machine learning segmentation model 31 available to operators for pulvinar segmentation, wholly set from previous training effected upstream by other players with the device 1.

[0092] In what follows, the modules are to be understood as functional entities rather than material, physically distinct, components. They can consequently be embodied either as grouped together in a same tangible and concrete component, or distributed into several such components. Also, each of those modules is possibly itself shared between at least two physical components. In addition, the modules are implemented in hardware, software, firmware, or any mixed form thereof as well. They are preferably embodied within at least one processor of the device 1 or of the device 2 or of the device 3.

[0093] The device 1 comprises a module 11 configured to receive as input the training dataset (i.e.; from a database 10) or to receive the input images required to construct said training dataset. In that case, module 11 may be configured to receive, for each subject of a plurality of subjects, at least one ground-truth image 52 and at least one FLAIR image 51. The at least one ground-truth image 52 and at least one FLAIR image 51 may be paired.

[0094] According to the present invention, "paired images" typically refer to a set of two different types of images that are acquired on a same subject (e.g., of said plurality of subjects) and represent a same structure (i.e., images of the brain wherein the pulvinar structure is observable or at least comprised even if not clearly visible).

[0095] According to the invention, the at least one ground-truth image 52 is a T1 mapping image, which is an image that may be acquired using magnetic resonance imaging (MRI) scan. T1 mapping images include a quantitative measure of the relaxation times of tissues within the imaged region. T1 relaxation time, is a fundamental property of tissues that determines the rate at which the nuclear magnetization of protons within those tissues returns to their equilibrium state after being perturbed by an MRI scanner's magnetic field. T1 mapping images may be obtained using the MP2RAGE sequence. It is a specific sequence used to acquire T1 mapping images with an improved contrast, an improved resolution and a reduced sensitivity to artifacts compared to traditional T1 mapping sequences. MP2RAGE stands for "Magnetization-Prepared 2 Rapid Acquisition Gradient Echo". The MP2RAGE sequence starts with a magnetization preparation step. This involves applying a series of radiofrequency (RF) pulses and magnetic gradients to manipulate the protons in the tissues. The purpose of this preparation is to saturate or nullify the magnetization in certain tissues, including fat, so that the resulting image will have improved contrast. After the magnetization preparation, the sequence employs a rapid 2D gradient echo imaging sequence to acquire data. This involves the application of RF pulses and magnetic gradients to generate signals from the tissues. The data acquisition process is typically repeated several times with different multiple inversion times. The multiple datasets collected with different inversion times are used to calculate the T1 relaxation times for each voxel in the image. This calculation is based on the exponential recovery of magnetization following the magnetization preparation.

[0096] Additionally, module 11 may be configured to receive the untuned machine learning segmentation model 50 (i.e., the machine learning segmentation model comprising initialization parameters), stored in one or more local or remote database(s) 10.

[0097] The database 10 can take the form of storage resources available from any kind of appropriate storage means, which can be notably a RAM or an EEPROM (Electrically- Erasable Programmable Read-Only Memory) such as a Flash memory, possibly within an SSD (Solid-State Disk).

[0098] The device 1 further comprises optionally a module 12 for preprocessing the at least one ground-truth image 52 and at least one FLAIR image 51. For instance, the atleast one ground-truth image 52 and at least one FLAIR image 51 may be cropped to focus on the signal within the brain and re-sized. The at least one ground-truth image 52 and at least one FLAIR image 51 may also be rescaled within a predefined range.

[0099] The device 1 may further comprise a module 13 configured to co-register the at least one ground-truth image 52 and at least one FLAIR image 51 in a training reference space. The training reference space may be the space of one of the at least one groundtruth image 52 or the space of one of the at least one FLAIR image 51, or an entirely new space. Additionally, the at least one ground-truth image 52 and at least one FLAIR image 51 may be first co-registered in the training reference space and then co-registered in a standard space such as the MNI (Montreal Neurological Institute) space.

[0100] The co-registration in the training reference space may be performed by first determining a registration matrix (i.e., transformation matrix) between the at least one ground truth image space and the training reference space. In other words, the transformation parameters (translation, rotation, scaling, and shearing) needed to align each subject's ground truth image to the training reference space are computed. These transformation parameters collectively form an affine registration matrix. In a second step, the registration matrix between the at least one FLAIR image space and the at least one ground truth image space is determined for each subject. Afterwards, the two registration matrices are combined so that the at least one FLAIR image can be registered into the training reference space. To that end, trilinear interpolation may be used.

[0101] Alternatively, the device 1 may further be configured to receive at least one Tl- weighted image from the same subject from which the at least one ground-truth image 52 and the at least one FLAIR image 51 were obtained.

[0102] According to the invention, T 1 - weighted images may be obtained using magnetic resonance imaging (MRI) and a specific MRI pulse sequence to highlight differences in tissue characteristics based on their T1 relaxation times. For instance, the Tl-weighted images may be obtained using gradient echo sequence, that is characterized by the use of gradients and radiofrequency (RF) pulses to create images with specific contrast properties. Advantageously, Tl-weighted images have the most anatomical information that may be used for registration. Therefore, using Tl-weighted images thus allows toimprove the quality of the intrasubject rigid registration between the at least one FLAIR image 51 and the at least one ground truth image 52.

[0103] Module 13 may then be configured to co-register the at least one Tl-weighted image and at least one FLAIR image 51 in the training reference space. In other words, module 13 may be configured to determine the registration matrix between the Tl- weighted image space and the training reference space. Then, the at least one ground truth image may be co-registered with the at least one Tl-weighted image. In other words, the registration matrix between the at least one ground truth image space and the at least one Tl-weighted image space may be determined. By combining the registration matrix between the at least one ground truth image space and the at least one Tl-weighted image space and the registration matrix between the at least one ground truth image space and the training reference space, it is possible to obtain the registration matrix of the at least one ground truth image 52 in the training reference space.

[0104] Additionally, the registration matrix between the training reference space and the standard space may be determined in order to co-register the at least one Tl-weighted image and at least one FLAIR image 51 in the standard space.

[0105] The device 1 may further comprise a module 14 configured to segment the pulvinar in the at least one ground-truth image 52 so to obtain at least one pulvinar segmentation mask (i.e., tridimensional matrix). The segmentation step may be performed manually by a trained expert, such as a physician or radiologist, that visually identifies and traces the boundaries of the pulvinar to create a binary mask, where the pulvinar is labeled as foreground and the rest as background. Alternatively, the segmentation may be performed automatically using any technique known from the skilled man in the art.

[0106] The segmentation of the pulvinar may be performed before the co-registration on the at least one ground-truth image 52 (e.g., on the T1 mapping images) or after coregistration. Additionally, the device 1 may be configured to directly receive the segmented ground-truth images 52 as input.

[0107] If the segmentation is performed before the co-registration step, the obtained pulvinar segmentation mask may also be co-registered in the training reference space. To that end, nearest neighbor interpolation may be used.

[0108] The device may comprise a module 15 for the generation of the training dataset from the received at least one FLAIR image 51 and the pulvinar segmentation mask obtained from module 14. The generated training dataset may comprise multiple training samples, each training sample associated to a different subject. A training sample may comprise the pulvinar segmentation mask, previously obtained on the ground truth image captured on a subject, and at least one FLAIR image captured on the same subject. Preferably, the pulvinar segmentation mask is co-registered with the at least one FLAIR image.

[0109] The device 1 further comprises a module 16 configured to train the segmentation model 50 using the training dataset constructed by the module 15 (or received by module 11).

[0110] The segmentation model 50, trained in module 16, is a neural network and may be a convolutional neural network (CNN) or an attention based neural network, such as a transformer, or any neural network for which the last layer outputs an activate map that has the size of a 3D image. The pulvinar segmentation mask obtained is a 3D binary image (i.e., tridimensional matrix).

[0111] In one example the segmentation model 50 is based on an architecture of a 3D U-Net model, such as illustrated on Figure 7. The 3D U-Net architecture, which is a CNN, is particularly well-suited for tasks where the input image needs to be segmented into distinct regions or objects. The 3D U-Net architecture consists of two main parts: the contracting path (encoder) and the expansive path (decoder). The contracting path captures the contextual information of the input image through a series of convolutional and pooling layers. This part of the network reduces the spatial dimensions of the input while extracting high-level features. The expansive path, also known as the decoder, aims to recover the spatial information and produce a segmentation map. It uses a series of up- sampling and convolutional layers to gradually increase the spatial resolution. Skip connections are incorporated between the contracting and expansive paths to enable the network to retain and merge high-resolution features with the up-sampled information. The skip connections are a key feature of the 3D U-Net architecture, allowing the network to capture both local and global context. They enable the precise localization of objects and help address the problem of information loss during the down-sampling process.

[0112] During the training, a loss function may be calculated. The loss function may comprise (or be equal) to a sum of a smooth dice loss and a Binary Cross Entropy with logits, such as the following:

[0113] iog(i - yt))

[0114] Where \PRED A GT\ is the number of overlapping voxels (e.g., identical subpart of the T1 mapping image and pulvinar segmentation mask) between the predicted pulvinar segmentation mask and the ground truth image, \PRED\ is the number of voxels in the predicted pulvinar segmentation mask and I GT\ the number of voxels in the ground truth image. N is the number of voxels, and for each z voxel, yt is the network prediction which is the probability to be or not part of pulvinar, and yt is the voxel value in the ground truth image (0 if not in pulvinar and 1 if in pulvinar).

[0115] Once the training completed, the module 14 is configured to output the trained ML segmentation model 31, notably the trained parameters (i.e., trained weights) for the trained ML segmentation model 31. The trained segmentation model 31 may then by stored in one or more local or remote database(s) 10. The latter can take the form of storage resources available from any kind of appropriate storage means, which can be notably a RAM or an EEPROM (Electrically-Erasable Programmable Read-Only Memory) such as a Flash memory, possibly within an SSD (Solid-State Disk).

[0116] In its automatic actions, the device 1 may for example execute the following process (Figure 2):- receiving a plurality of image pairs (e.g., from a plurality of subjects), each pair of images comprising at least one ground-truth image, said ground-truth image being a T1 -weighted image, and at least one FLAIR image (step 41);- optionally preprocessing the at least one ground-truth image and the at least on FLAIR image (step 42);- co-registering the at least one ground-truth image and the at least one FLAIR image in a training reference space (step 43);- segmenting the pulvinar in the at least one ground-truth image so as to obtain at least one pulvinar segmentation mask (step 44);- generating a training dataset comprising multiple training samples associated to a plurality of subjects, each training sample comprising the co-registered pulvinar segmentation mask and the at least one FLAIR image from the same subject (step 45);- training a convolutional neural network with the generated training dataset so as to obtain, as output, a prediction of a pulvinar segmentation mask for an input FLAIR image of a patient (step 46).

[0117] The present invention also relates to the device 2 for diagnosis of Status Epilepticus (SE) in a patient. The device 2 will be described in reference to a particular function embodiment as illustrated in Figure 3.

[0118] The device 2 is adapted to receive as input the trained machine learning segmentation model 31 (i.e., machine learning segmentation model 50 with the trained parameters) and at least one FLAIR image 32 of the brain, at least one BO image 33 of the brain, and at least one B1000 image 34 of the brain from a same patient.

[0119] In diffusion- weighted imaging (DWI), B-values refer to the different levels of diffusion weighting applied to the MRI sequence. B-values are used to probe the diffusion of water molecules in tissues and can provide information about tissue microstructure. BO and B1000 are two commonly used B-values in DWI.

[0120] Therefore, a B0 image is a diffusion-weighted image where a B-value of 0 s / mm2has been selected for the acquisition and a Bl 000 image is a diffusion- weighted image where a B-value of 1000 s / mm2has been selected for the acquisition.

[0121] The device 2 is adapted to segment the pulvinar region in either or both the B0 and B1000 DWI images and to perform a measurement on said segmented region. The measurement is then compared to a predefined criterion and the device 2 is adapted to provide as an output a diagnosis of the patient as being affected with SE or not. The diagnosis may further comprise a severity of SE for the patient.

[0122] To that end, the device 2 comprises a module 17 for receiving the trained machine learning segmentation model 31 and the at least one FLAIR image 32 of the brain, the at least one B0 image of the brain, and at least one B1000 image of the brain, that may be stored in one or more local or remote database(s) 10. The latter can take the form ofstorage resources available from any kind of appropriate storage means, which can be notably a RAM or an EEPROM (Electrically-Erasable Programmable Read-Only Memory) such as a Flash memory, possibly within an SSD (Solid-State Disk). In advantageous embodiments, the trained machine learning segmentation model 31 and all its training parameters have been previously generated by a system including the device 1 for training. Alternatively, the trained machine learning segmentation model 31 and its training parameters are received from a communication network.

[0123] The device 2 further comprises optionally a module 18 for preprocessing the at least one FLAIR image 32, the at least one BO image 33, and at least one B1000 image 34. For instance, the at least one FLAIR image 32, the at least one BO image 33, and at least one Bl 000 image 34 may be cropped to focus on the signal within the brain and resized. They may also be rescaled within a predefined range.

[0124] The device 2 may further comprise a module 19 configured to feed the at least one FLAIR image 32 to the trained segmentation model 31 to obtain a pulvinar segmentation mask for said at least one FLAIR image 32.

[0125] The device 2 may also comprise a module 20 configured to co-register the at least one FLAIR image 32, the at least one B0 image 33 and the at least one B1000 image 34 in a reference space. The reference space may be the space of one of the at least one FLAIR image 32 or the space of one of the at least one B0 image 33, or the space of one of the at least one B1000 image 34 or an entirely new space.

[0126] The co-registration may be performed by first determining the registration matrix (i.e., transformation matrix) between the at least one FLAIR image space and the reference space. In other words, the transformation parameters (translation, rotation, scaling, and shearing) needed to align each subject's ground truth image to the training reference space are computed. These transformation parameters collectively form an affine registration matrix.

[0127] In a second step, the registration matrix between the B0 image space and the FLAIR image space may be determined. In the same manner, the registration matrix between the Bl 000 image space and the FLAIR image space may be determined. Afterwards, the registration matrices may be combined so that the at least one B0image 33 and the at least one B1000 image 34 can be registered into the reference space. To that end, trilinear interpolation may be used.

[0128] The device 2 may also comprise a module 21 for applying the obtained pulvinar segmentation mask to the at least one BO image and to the at least one B1000 image in order to obtain a BO-segmented region (i.e., volume) and a BlOOO-segmented region (i.e., volume).

[0129] The device 2 may also comprise a module 22 for calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one BO image and / or for calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image. The term voxels intensity refers to the level of signal intensity associated with each individual voxel in the 3D MRI images.

[0130] The device 2 may also comprise a module 23 for comparing said ADC and / or said distribution of voxels intensity to a predefined criterion.

[0131] The predefined criterion may be an ADC calculated on the segmented pulvinar region of at least one B0 image obtained on a healthy control (i.e., contralateral pulvinar of the same subject or different healthy subject) and / or a distribution of voxels intensity calculated on the segmented pulvinar region of at least one B1000 image obtained on a healthy control. The subject may be diagnosed with SE when the patient ADC is lower than said ADC calculated on the healthy control and / or when the patient distribution of voxels intensity is shifted towards higher intensity values than said distribution of voxels intensity calculated on the healthy control.

[0132] The difference of distribution between the patient and the healthy control (i.e., quantification of said shift towards higher intensity values) could either be based on descriptive statistics such as median mean and standard deviation, based on the identification of several components of the signal using Gaussian Mixture Model and Bayesian Gaussian Mixture Models, or based on the sampling of a subset of voxels in the segmented pulvinar region. This sampling of a subset of voxels may be done for example by thresholding the Xthdecile of the voxels based on the B 1000 value and looking then at their ADC, or using a Euclidian distance map from lateral ventricles to extract the diffusion value of the closest voxels to the lateral ventricles.

[0133] In examples, anomaly detection algorithms derived from machine learning techniques, such as Gaussian Mixture Models or Bayesian Gaussian Mixture Models, are applied to the distribution of diffusion (i.e., ADC) values in pulvinar voxels (i.e., voxels in the segmented pulvinar region). This helps detect specific peaks in patients with SE compared to healthy controls or a reference distribution (e.g. ADC distribution).

[0134] The device 2 may interact with a user interface 82, via which information can be entered and retrieved by a user. The user interface 82 includes any means appropriate for entering or retrieving data, information or instructions, notably visual, tactile and / or audio capacities that can encompass any or several of the following means as well known by a person skilled in the art: a screen, a keyboard, a trackball, a touchpad, a touchscreen, a loudspeaker, a voice recognition system.

[0135] In its automatic actions, the device 2 may for example execute the following process (Figure 4):- receiving the at least one FLAIR image 32 of the brain, at least one BO image 33 of the brain, and at least one B1000 image 34 of the brain (step 61);- optionally preprocessing the at least one FLAIR image 32, at least one BO image 33 and at least one B1000 image 34 (step 62);- feeding the preprocessed at least one FLAIR image 32 to said trained machine learning segmentation model 31 so to generate at least one pulvinar segmentation mask (step 63);- co-registering the at least one FLAIR image 32, the at least one BO image 33 and the at least one B1000 image 34 in a reference space (step 64);- applying the obtained pulvinar segmentation mask to the at least one BO image and to the at least one B1000 image so as to obtain at least one BO-segmented region and at least one BlOOO-segmented region (step 65);- calculating an ADC (Apparent Diffusion Coefficient) on at least one BO-segmented region and / or calculating a distribution of voxels intensity on the and at least one BlOOO-segmented region (step 66);- comparing said ADC and / or said distribution of voxels intensity to a predefined criterion (step 67).

[0136] The present invention also relates to the device 3 for assessing the severity 62 of Status Epilepticus (SE) in a patient using a trained machine learning segmentation model for pulvinar segmentation obtained from the device 1. The device 3 will be described in reference to a particular function embodiment as illustrated in Figure 5.

[0137] The device 3 is adapted to receive as input the trained machine learning segmentation model 31 (i.e., machine learning segmentation model 50 with the trained parameters) and at least one FLAIR image 32 of the brain, at least one B0 image 33 of the brain, and at least one B1000 image 34 of the brain from a same patient.

[0138] The device 3 is adapted to segment the pulvinar region in either or both the B0 and B1000 DWI images and to perform a measurement on said segmented region. The measurement is then compared to a predefined criterion and the device 3 is adapted to provide as an output the severity 62 of Status Epilepticus (SE) for the patient.

[0139] The device 3 may comprise a module 24 for receiving the trained machine learning segmentation model 31 and the at least one FLAIR image 32 of the brain, the at least one B0 image of the brain, and at least one B1000 image of the brain, that may be stored in one or more local or remote database(s) 10. Module 24 may implement the same functionalities as module 17 of device 2.

[0140] The device 3 may further comprise modules 25, 26, 27, 28, 29 and 30 that have similar functions as modules 18, 19, 20, 21, 22 and 23 of device 2 respectively.

[0141] The device 3 may further comprise at least one output configured to output the severity 62 of SE. The severity of SE may be a probability value, a percentage, a range of values, or categorized into specific levels such as mild, moderate, severe, and critical, each indicating varying degrees of severity of SE.

[0142] The device 3 may also interact with a user interface 82, via which information can be entered and retrieved by a user.

[0143] In its automatic actions, the device 3 may for example execute the following process (Figure 6):receiving from said patient at least one FLAIR image of the brain, at least one BO image of the brain, and at least one B1000 image of the brain (step 71); feeding the at least one FLAIR image to said trained segmentation model to obtain at least one pulvinar segmentation mask (step 73); co-registering the at least one FLAIR image, the at least one BO image and the at least one B1000 image in a reference space (step 74); applying the at least one pulvinar segmentation mask to the at least one BO image and to the at least one Bl 000 image (step 75); calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one BO image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image (step 76); comparing said ADC and / or said distribution of voxels intensity to a predefined criterion (step 77); outputting the severity 62 of the SE based on said comparison of said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion.

[0144] According to one embodiment, the SE is considered severe if the ADC is decreased as compared to an ADC predefined criterion. Such an assessment may enable a medical professional to confirm the diagnosis of SE while also evaluating its severity. A significant decrease in the ADC may be indicative of ischemic phenomena occurring in the brain, which can precipitate seizures, including SE. Therefore, the severity 62 of the SE may be invaluable for guiding therapeutic interventions, determining prognosis, and monitoring treatment response. In essence, the ability to accurately assess an ADC variation confirms the diagnosis of SE and provides as well a quantitative measure of its severity, thereby enhancing clinical decision-making and patient management.

[0145] A particular apparatus may embody the device 1 as well as the devices 2 and / or 3 described above. It corresponds for example to a workstation, a laptop, a tablet, a smartphone, or a head- mounted display (HMD).

[0146] That apparatus is suited to generation of segmentation mask and to related Machine Learning training. It comprises the following elements, connected to each other by a bus of addresses and data that also transports a clock signal:- a microprocessor (or CPU);- a graphics card comprising several Graphical Processing Units (or GPUs) and a Graphical Random Access Memory (GRAM); the GPUs are quite suited to image processing, due to their highly parallel structure;- a non-volatile memory of ROM type;- a RAM;- one or several I / O (Input / Output) devices such as for example a keyboard, a mouse, a trackball, a webcam; other modes for introduction of commands such as for example vocal recognition are also possible;- a power source; and- a radiofrequency unit.

[0147] According to a variant, the power supply is external to the apparatus.

[0148] The apparatus also comprises a display device of display screen type directly connected to the graphics card to display synthesized images calculated and composed in the graphics card. According to a variant, a display device is external to the apparatus and is connected thereto by a cable or wirelessly for transmitting the display signals. The apparatus, for example through the graphics card, comprises an interface for transmission or connection adapted to transmit a display signal to an external display means such as for example an LCD or plasma screen or a video -projector. In this respect, the RF unit can be used for wireless transmissions.

[0149] It is noted that the word "register" used hereinafter in the description of memories can designate in each of the memories mentioned, a memory zone of low capacity (some binary data) as well as a memory zone of large capacity (enabling a whole program to be stored or all or part of the data representative of data calculated or to be displayed). Also, the registers represented for the RAM and the GRAM can be arranged and constituted in any manner, and each of them does not necessarily correspond to adjacent memorylocations and can be distributed otherwise (which covers notably the situation in which one register includes several smaller registers).

[0150] When switched-on, the microprocessor loads and executes the instructions of the program contained in the RAM.

[0151] As will be understood by a skilled person, the presence of the graphics card is not mandatory, and can be replaced with entire CPU processing and / or simpler visualization implementations.

[0152] In variant modes, the apparatus may include only the functionalities of the device 1, and not the learning capacities of the device 2. In addition, the device 1 and / or the device 2 and / or the device 3 may be implemented differently than a standalone software, and an apparatus or set of apparatus comprising only parts of the apparatus may be exploited through an API call or via a cloud interface.

[0153] According to another aspect, the invention relates to a device for diagnosis of Status Epilepticus (SE) in a patient using a trained machine learning segmentation model for pulvinar segmentation obtained from a device (such as described above), said device comprising: at least one input configured to receive from said patient at least one FLAIR image of the brain, at least one B0 image of the brain, and at least one B1000 image of the brain; at least one processor configured to:• feed the at least one FLAIR image to the trained machine learning segmentation model to obtain at least one pulvinar segmentation mask;• co-register the at least one FLAIR image, the at least one B0 image and the at least one B 1000 image in a reference space;• apply the obtained pulvinar segmentation mask to the at least one B0 image (i.e., so as to obtain the segmented pulvinar region of the at least one B0 image) and to the at least one B1000 image (i.e., so as to obtain a the segmented pulvinar region of the at least one B 1000 image);• calculate an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image, and / or calculate a distributionof voxels intensity on the segmented pulvinar region of the at least one Bl 000 image;• compare said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion; and at least one output configured to output the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

[0154] According to another aspect, the invention further relates to a method of treating or preventing Status Epilepticus (SE) in a patient in need thereof, using the trained machine learning segmentation model for pulvinar segmentation obtained from the device described, said method comprising: receiving from said patient at least one FLAIR image of the brain, at least one BO image of the brain, and at least one B 1000 image of the brain; feeding the at least one FLAIR image to said trained segmentation model to obtain at least one pulvinar segmentation mask; co-registering the at least one FLAIR image, the at least one B0 image and the at least one B1000 image in a reference space; applying the at least one pulvinar segmentation mask to the at least one B0 image and to the at least one Bl 000 image; calculating an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image; comparing said ADC and / or said distribution of voxels intensity to a predefined criterion; outputting the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion; and treating or preventing SE in the diagnosed patient.

[0155] According to one embodiment, the step of treating or preventing SE in the diagnosed patient comprises the administration of a treatment to said patient. According to one embodiment, the treatment is to be administered at a therapeutically effective amount to said patient.

[0156] According to one embodiment, said treatment may for instance be Benzodiazepines, levetiracetam, sodium valproate, fosphenytoin, phenobarbital or propofol.

[0157] According to one embodiment, said treatment is selected from the group comprising or consisting of levetiracetam (CAS number 102767-28-2), sodium valproate (CAS number 1069-66-5), fosphenytoin disodium (CAS number 92134-98-0), phenobarbital (CAS number 50-06-6), propofol (CAS number 2078-54-8), lacosamide (CAS number 175481-36-4), ketamine (CAS number 6740-88-1) and benzodiazepines such as, for example clonazepam (CAS number 1622-61-3), diazepam (CAS number 439- 14-5), lorazepam (CAS number 846-49-1), clobazepam (CAS number 22316-47-8) or midazolam (CAS number 59467-70-8).

[0158] According to one embodiment, said treatment is a benzodiazepine. According to one embodiment, said treatment is levetiracetam (CAS number 102767-28-2). According to one embodiment, said treatment is sodium valproate (CAS number 1069-66-5). According to one embodiment, said treatment is fosphenytoin disodium (CAS number 92134-98-0). According to one embodiment, said treatment is phenobarbital (CAS number 50-06-6). According to one embodiment, said treatment is lacosamide (CAS number 175481-36-4). According to one embodiment, said treatment is ketamine (CAS number 6740-88-1). According to one embodiment, said treatment is clonazepam (CAS number 22316-47-8). According to one embodiment, said treatment is diazepam (CAS number 439-14-5). According to one embodiment, said treatment is lorazepam (CAS number 846-49-1). According to one embodiment, said treatment is clobazepam (CAS number 22316-47-8). According to one embodiment, said treatment is midazolam (CAS number 59467-70-8).

[0159] According to one embodiment, said patient is affected with a disease or condition at risk of status epilepticus. Examples of diseases or conditions at risk of status epilepticus include, without limitation, epilepsy, encephalitis, meningoencephalitis, intracranial hemorrhage, ischemic stroke, Traumatic Brain Injury (TBI), brain tumors, neurodegenerative diseases, Creutzfeldt Jacob disease.

[0160] According to one embodiment, said patient is affected with a disease or condition selected from the group comprising or consisting of epilepsy, encephalitis, meningoencephalitis, intracranial hemorrhage, ischemic stroke, Traumatic Brain Injury (TBI), brain tumors, neurodegenerative diseases, and Creutzfeldt Jacob disease.

[0161] According to one embodiment, said patient is affected with epilepsy.

[0162] According to one embodiment, said patient is affected with encephalitis.

[0163] According to one embodiment, said patient is affected with meningoencephalitis .

[0164] According to one embodiment, said patient is affected with intracranial hemorrhage.

[0165] According to one embodiment, said patient is affected with ischemic stroke.

[0166] According to one embodiment, said patient is affected with Traumatic Brain Injury (TBI).

[0167] According to one embodiment, said patient is affected with a brain tumor, such as , for example glioblastoma multiforme (GBM), aanaplastic astrocytoma, meningioma, oligodendroglioma, metastatic brain tumor, ganglioglioma, dysembryoplastic neuroepithelial tumor (DNET), ependymoma, medulloblastoma, or primary central nervous system lymphoma . As disclosed herein a brain tumor is an abnormal growth of cells within or around the brain which can be either benign (non cancerous) or malignant (cancerous). Brain tumors may disrupt normal brain function by compressing or invading surrounding tissue, leading to symptoms such as headaches, seizures, cognitive changes or neurological deficits. Brain tumors my originate from the brain (primary brain tumor) or have spread from other parts of the body (metastatic tumor).

[0168] According to one embodiment, said patient is affected with a neurodegenerative disease, such as, for example, Alzheimer’s disease, Frontotemporal Dementia, Lewy Body Dementia, Corticobasal Degeneration, Creutzfeldt-Jacob disease, Progressive Myoclonic Epilepsies, Huntington’s disease, Multiple System Atrophy, Niemann-Pick Disease Type C, or Subacute Sclerosing Panencephalitis. As disclosed herein, a neurodegenerative disease is a disorder characterized by the progressive loss of structure or function of neurons, leading to gradual neurological decline and symptoms such as memory loss, impaired motor skills, and cognitive dysfunction.

[0169] According to one embodiment, said patient is affected with Creutzfeldt- Jacob disease.

[0170] According to one embodiment, the SE presents with convulsions. According to another embodiment, the SE presents no motor manifestations.

[0171] It will be understood by the skilled artisan in the art that the methods and devices as described herein may present the following advantages.

[0172] The methods and devices as presented herein may allow to diagnose SE when SE presents with convulsions, but also when SE presents no motor manifestations.

[0173] The methods and devices as presented herein may allow rapid management of the patient once the SE is diagnosed.EXAMPLES

[0174] The present invention is further illustrated by the following example.Materials and Methods

[0175] Three already acquired MRI datasets of Healthy Controls (HC) and Multiple Sclerosis (MS) patients were used to build the cohorts from this example.

[0176] The first cohort (named ONSTIM) comprises 28 MS patients (19 women and 9 men) of age 34.3 on average, with a mean Expanded Disability Status Scale (EDSS) of 0.8, and 10 sex and age-matched Healthy controls (HC). The ten healthy controls and 12randomly chosen MS patients were selected from this cohort.

[0177] The second cohort (named ENERGYSEP) comprises 24 MS patients (15 women and 9 men) of age 36.5 on average, with a mean EDSS of 2.05, and 12 HC (5 women and 7 men) of age 35.1 on average. 11 HC and 18 MS patients were randomly selected from this second cohort.

[0178] The third cohort (named SEPBIOPROGRESS) comprises 75 MS patients (59 women 16 men) of age 38.7 in average, with mean EDSS 2.23. 14 MS patients were randomly selected from this last cohort.

[0179] The two first cohorts may be used to train the segmentation algorithm. The addition of the two first cohorts comprises 51 subjects (30 MS patients and 21 HC). Those 51 subjects were split in 49 subjects for training and 2 for validation. The last cohort may be used for testing only, representing 14 subjects (all MS patients).

[0180] All HC and MS patients that were randomly selected from the previous cohorts underwent MRI acquisitions with sequences FLAIR (e.g., T2FLAIR), MP2RAGE, gradient echo Tl, and Diffusion Weighted Imaging (DWI). B0 and B1000 DWI images, FLAIR images, Tl-weighted images and Tl mapping images were acquired on a 3Testla (T) Siemens Prisma machine. Tlmapping images obtained using MP2RAGE sequence were acquired on the same 3T Siemens Prisma machine for the first cohort and the third cohort, and on a 7T Siemens Magnetom machine for the second cohort.

[0181] For all subjects, the Tlmapping image was used to segment the pulvinars. Online available WMnull atlas were used to seek for anatomical repairs that may help for the segmentation. The pulvinar was segmented manually using itk-SNAP free, open-source, multi-platform software application.

[0182] At the end of this step, manually segmented binary pulvinar masks corresponding to ground truth in native space were obtained.

[0183] For each image from the first cohort and the second cohort, an affine registration into a standard space was performed. For that, Advanced Normalization Tools was used. Each subject’s FLAIR image was then co-registered with his Tl-weighted image. Bycombining the two precedent registration matrices, a registration from the native FLAIR space into the standard space using trilinear interpolation was performed.

[0184] The subject’s Tlmapping image was also co-registered with the subject’s Tl- weighted image, and by combining registration matrix obtained from the Tlmapping image to the Tl-weighted image and from the Tl-weighted image to the standard space, it was possible to register the pulvinar manual segmentation to the standard space using nearest neighbor interpolation.

[0185] At the end of this step, it is possible to obtain, for all subjects of the first cohort and the second cohort, a dataset comprising the FLAIR images and the dataset comprising the ground truth images in standard space. To limit memory usage, the images were cropped in the standard space, so as to obtain images of 146x182x156 voxels in the standard space, with voxel size corresponding to 1 mm3. This template registration is known to help segmentation network to learn fixed anatomical features.

[0186] Intensities were then rescaled in a reproductible manner between the FLAIR images and ground truth images. In the standard space, a brain mask was binarized in order to create an exclusively parenchymal mask of interest in which the intensities may be rescaled in a comparable manner, excluding the skull. Indeed, the skull is highly hyperintense on FLAIR images and this hyperintensity would perturbate the rescaling process for the parenchyma. Into this fix eroded brain mask, the intensities from [2.5%; 97.5%] of the range of their native values to [ -1;1] were rescaled. This helped to make parenchymal contrast similar between different FLAIR images.

[0187] The paired FLAIR images and ground truth images were then converted into pytorch tensors, and split randomly into 49 patients for training and 2 for validation.

[0188] Training was run on Google COLAB ® pro+, allowing to access NVIDIA® K80, T4 and P100 Graphical Cards, with maximum GB RAM accessible. Batch size was limited to 1 due to GPU constraints. Torchio library was used to add data augmentation so that the network would learn a lot of variability from the training set and would have a better generalizability.

[0189] A standard 3D U-net was used and the sum of the smooth dice loss and BinaryCross Entropy (BCE) with logits was used as loss function.

[0190] The network was trained during 100 epochs. Learning rate was initially 0.0001 and was decreased based on the validation (it decreased of factor 2 each time validation loss reaches a plateau), by using torchio scheduler tool.

[0191] The trained network was applied to the 14 FLAIR images from the third cohort. The preprocessing for testing FLAIR images was the same as previously described (cropping and rescaling). Moreover, affine registration between the FLAIR images and standard space was performed without using T1 -weighted images. As the output is a probability, it was applied to a sigmoid, and binarized with a threshold of 0.5. All predictions outside of the fix eroded brain mask were put to zero as they cannot correspond to pulvinar. The binary predicted pulvinar mask was then registered back to the native FLAIR space using nearest neighbor interpolation.

[0192] The ground truth of the testing set was manually segmented on T1 mapping images on MP2RAGE space. A rigid registration towards a same subject’s FLAIR image was performed, by using nearest neighbor interpolation with Ants.

[0193] To compare the trained machine learning segmentation model pulvinar segmentation masks predictions with ground truth in the FLAIR space, a dice score was used. The binary mask from the ground truth and from the segmentation model was labialized using a connected component analysis performed on Python 3 with scikt- image library. The left side was then given the arbitrary value of 1 and the right side was given the arbitrary value of 2 in the pulvinar segmentation map.

[0194] Dwi2adc command from MRTRIX was used to extract the ADC and the Trace image from DWI images. A rigid registration was then performed for each testing patient from his Trace image to his FLAIR image with Ants, and the inverse transform was used to export binary pulvinar masks (either from ground truth or from the PulviNet) to the ADC map using nearest neighbor interpolation. The mean ADC value in right and left pulvinar were then extracted with nibabel.

[0195] All statistics were done on Python 3 using scipy. stats library, and setting type Ierror rate to 0.05. Wilcoxon paired signed rank test was used to test a difference between pulvinar volumes from the manual segmentation and from the PulviNet, to test a difference between ADC value extracted by the PulviNet and by manual segmentation, and also to test if there was a difference between right and left pulvinar’ s ADC value. Spearman correlation was used to see if predicted and manually segmented pulvinar volumes were correlated, and to see if ADC values extracted from PulviNet and from manual segmentation were correlated, for the right and for the left pulvinar. All correlation graphs and boxplot are made with seaborn library (https: / / seabom.pydata.org / ) in Python 3.Results

[0196] On the 14 patients in the third cohort, a dice between segmented ground truth and the trained pulvinar segmentation model prediction of 0.715 (sd = 0.098) is obtained, as illustrated on Figure 9. Examples of segmentation output are shown on Figure 8. In the testing set, on FLAIR space, the average pulvinar volume was 2327,1 mm3 (sd = 1059.9) in ground truth vs 2272.4 mm3 (sd = 548.5) for the trained machine learning segmentation model for pulvinar segmentation prediction, which corresponds to a non- significative relative decrease of 2.36% (p = 1.0). A significant correlation was obtained between pulvinar volume from the trained machine learning segmentation model prediction and from manual segmentation for the 14 subjects of the third cohort (r = 0.75, p = 0.002), as shown in Figure 10.

[0197] After right- left labialization and registration in diffusion space, an excellent correlation between the mean pulvinar’ s ADC value is extracted for each test subject with the trained machine learning segmentation model and manual segmentation for the right pulvinar (r = 0.96, p = 5.08 .10-8), as well as for the left pulvinar (r = 0.94, p = 4.4 .10- 7), as shown on Figure 12 and Figure 13. No significant difference between ADC extracted from right and from left pulvinar within the ground truth was found (relative difference of -1.46%, sd = 3.59%, p = 0.24), as shown on Figure 9, and neither between ADC extracted from right and from left pulvinar with the trained machine learning segmentation model (relative difference of 0.22%, sd = 4.38%, p = 0.22).

[0198] To conclude, the contribution of this example on the clinical front is twofold: confirming the diagnosis of status epilepticus and assessing its severity. For instance, the severity of SE may be assessed by measuring the intensity of ADC variation. Indeed, a big decrease in ADC is associated to ischemic phenomenon in the brain, which refer to a reduced blood flow and oxygen supply to a specific area of the brain, that can trigger seizures, including SE. Therefore, the device 3 of the present invention, by quantifying the decrease in ADC can then measure the magnitude of the ischemic phenomenon caused by SE in pulvinars.

[0199] Status epilepticus is a highly heterogeneous condition that can complicate various pathologies, either of systemic origin (intoxication, electrolyte imbalance) or cerebral (stroke, tumors, encephalitis, etc.). Diagnosis, while easy in the presence of convulsions, becomes challenging when there is a sudden neurological deficit without motor manifestations. The role of MRI is crucial in such cases for diagnosing conditions like stroke, migraine aura, neurofunctional disorders, and status epilepticus. The pulvinar sign is one of the key elements for the diagnosis of status, enabling appropriate urgent management of the patient. In patients with known status epilepticus, the presence of pulvinar sign indicates a prolonged duration of seizures, allowing for therapeutic adjustments.

Claims

CLAIMS1. Device (1) for obtaining a trained machine learning segmentation model (31) for pulvinar segmentation, said device comprising: at least one input configured to receive a plurality of image pairs, each pair of images comprising: o at least one ground-truth image (52), said ground-truth image being a T1 mapping image, and o at least one FLAIR (Fluid Attenuated Inversion Recovery) image (51), wherein the at least one ground-truth image and the at least one FLAIR image are obtained from a same subject of a plurality of subjects; at least one processor configured to: o co-register the at least one ground-truth image (52) and the at least one FLAIR image (51) in a training reference space; o segment the pulvinar in the at least one ground-truth image (52) so as to obtain at least one pulvinar segmentation mask; o generate a training dataset comprising multiple training samples associated to said plurality of subjects, each training sample comprising the at least one pulvinar segmentation mask and the at least one FLAIR image (51) from the same subject, wherein the at least one pulvinar segmentation mask and the at least one FLAIR image (51) are co-registered in said training reference space; o obtain said trained machine learning segmentation model (31) by training, using said training dataset, a convolutional neural network so as to obtain as output, a prediction of a pulvinar segmentation mask based on an input FLAIR image of a patient, at least one output configured to provide the obtained trained machine learning segmentation model (31).

2. The device (1) according to claim 1, wherein the at least one ground- truth image (52) and the at least one FLAIR image (51) are co-registered into a standard space.

3. The device (1) according to claim 1 or 2, wherein the co-registration of the at least one ground-truth image (52) and the at least one FLAIR image (51) in the training reference space is performed using trilinear interpolation.

4. The device (1) according to any of claims 1 to 3, wherein the convolutional neural network is a 3D U-net.

5. The device (1) according to any of claims 1 to 4, wherein, during the training, the at least one processor is further configured to calculate a loss function based on a sum of a smooth dice loss and a Binary Cross Entropy with logits.

6. A device (2) for diagnosis of Status Epilepticus (SE) in a patient using a trained machine learning segmentation model (31) for pulvinar segmentation obtained from a device (1) of any of claims 1 to 5, said device (2) comprising: at least one input configured to receive from said patient at least one FLAIR image (51) of the brain, at least one BO image (53) of the brain, and at least one B1000 image (34) of the brain; at least one processor configured to:• feed the at least one FLAIR image (32) to the trained segmentation model (31) to obtain a pulvinar segmentation mask;• co-register the at least one FLAIR image (32), the at least one BO image (33) and the at least one B1000 image (34) in a reference space;• apply the obtained pulvinar segmentation mask to the at least one BO image (33) and to the at least one B1000 image (34);• calculate an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one BO image, and / or calculate a distribution of voxels intensity on the segmented pulvinar region of the at least one B1000 image;• compare said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion, andat least one output configured to output the diagnosis of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

7. The device (2) according to claim 6, wherein comparing further comprises calculating said ADC predefined criterion as an ADC on the segmented pulvinar region of at least one BO image obtained on a healthy control and / or calculating said intensity predefined criterion as a distribution of voxels intensity on the segmented pulvinar region of at least one B1000 image obtained on a healthy control, the subject being diagnosed with SE when the patient ADC is lower than said ADC calculated on the healthy control and / or when the patient distribution of voxels intensity is shifted towards higher intensity values than said distribution of voxels intensity calculated on the healthy control.

8. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out a computer-implemented method of diagnosing Status Epilepticus (SE) in a patient using a trained machine learning segmentation model (31) for pulvinar segmentation obtained thanks to the device (1) for obtaining said trained machine learning segmentation model (31) for pulvinar segmentation, said method comprising: receiving (17) from said patient at least one FLAIR image (32) of the brain, at least one BO image (33) of the brain, and at least one B 1000 image (34) of the brain; feeding (19) the at least one FLAIR image (32) to said trained segmentation model (31) to obtain at least one pulvinar segmentation mask; co-registering (20) the at least one FLAIR image (32), the at least one BO image (33) and the at least one B1000 image (34) in a reference space; applying (21) the at least one pulvinar segmentation mask to the at least one BO image (33) and to the at least one B1000 image (34);calculating (22) an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one BO image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B 1000 image; comparing (23) said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion; and outputting the diagnosis (61) of the patient as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

9. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out a computer-implemented method of diagnosing Status Epilepticus (SE) in a patient using a trained machine learning segmentation model (31) for pulvinar segmentation obtained thanks to the device (1) for obtaining said trained machine learning segmentation model (31) for pulvinar segmentation, said method comprising: receiving (17) from said patient at least one FLAIR image (32) of the brain, at least one BO image of the brain (33), and at least one B 1000 image (34) of the brain; feeding (19) the at least one FLAIR image (32) to said trained segmentation model (31) to obtain at least one pulvinar segmentation mask; co-registering (20) the at least one FLAIR image (32), the at least one B0 image (33) and the at least one B1000 image (34) in a reference space; applying (21) the at least one pulvinar segmentation mask to the at least one B0 image (33) and to the at least one B1000 image (34); calculating (22) an ADC (Apparent Diffusion Coefficient) on the segmented pulvinar region of the at least one B0 image and / or calculating a distribution of voxels intensity on the segmented pulvinar region of the at least one B 1000 image;comparing (23) said ADC to an ADC predefined criterion and / or said distribution of voxels intensity to an intensity predefined criterion; and outputting the diagnosis (61 ) of the subject as being affected with SE when the ADC complies with said ADC predefined criterion and / or said distribution of voxels intensity complies with said intensity predefined criterion.

Images