Dementia diagnosis

Abstract

Diagnosis of dementia. A method for diagnosing a dementia subtype is described. The method includes acquiring brain imaging data relating to one or more patients, analyzing the data using a deep learning model, and classifying the one or more patients into a plurality of categories using the deep learning model, including a first category of patients having a first subtype of dementia and a second category of patients having a second subtype of dementia. The deep learning model has been trained using brain imaging data from a patient, the brain imaging data includes a first set of images showing evidence of temporal-apical hypometabolism and a second set of images showing evidence of hypometabolism in the brain in place of or in addition to temporal-and apical regions, where the first set of images is marked as associated with a first subtype of dementia, and where the second set of images is marked as associated with a second subtype of dementia. And the second set of images are marked as being related to a second subtype of dementia. Related systems and methods are also described.

Description

Technical Field

[0001] The present invention relates to a method for diagnosing a subtype of dementia, and in particular but not exclusively to a method for diagnosing a subtype of dementia by analysing brain images, in particular ( 18 A method for classifying a patient as having Alzheimer's disease (AD) or mixed dementia, such as AD and cerebrovascular disease (CVD, mixed dementia), based on F)-fluorodeoxyglucose-positron emission tomography (FDG-PET) images. Background Art

[0002] Alzheimer's disease (AD) can coexist with other brain pathologies that also contribute to cognitive decline and complicate the diagnosis and treatment of AD. AD is often associated with cerebrovascular disease (CVD), and the presence of these two pathologies has an additive effect on cognitive decline. CVD is associated with decreased cognitive performance and lowers the threshold for clinical manifestations of dementia in AD patients. The overlap between these two pathologies has led to the term "mixed dementia."

[0003] The coexistence and heterogeneity of pathologies in patients with dementia make it difficult to distinguish between typical AD and mixed pathologies. It has been reported that FDG-PET is superior to other neuroimaging techniques, such as magnetic resonance imaging (MRI), computed tomography (CT), and single photon emission computed tomography (SPECT), in distinguishing other pathologies from AD. However, the visual interpretation of FDG-PET images requires intensive training of experts and is very time-consuming. In addition, the visual interpretation of FDG-PET images of AD is subjective and depends on expert knowledge (Morbelli et al., J Alzheimers Dis. 2015; 44(3): 815-26), with the consistency of expert visual analysis and clinical diagnosis being approximately 90% (Tripathi et al., Neuroradiol J 2014; 27(1): 13-21).

[0004] In addition, visual interpretation of FDG-PET images may miss subtle hypometabolism (Jo et al., Front. Aging Neurosci. 2019; 11). This can be detected by semiquantitative image analysis, using scores or statistics calculated for voxel values ​​in the image. It is reported that the sensitivity and specificity of semiquantitative analysis are as high as 93% in distinguishing possible AD from age-matched controls (Herholz et al., Neuroimage. 2002 September; 17 (1): 302-16). However, it is not without limitations, because semiquantitative analysis may easily produce inaccurate results, especially when evaluating very small or adjacent areas in the brain (Sarikaya et al., J Nucl Med Tech 2018 December, 46 (4) 362-367).

[0005] Therefore, there is a need for improved methods of diagnosing AD, particularly methods of differentiating AD from mixed dementias, which do not suffer from the shortcomings of the prior art. Summary of the invention

[0006] In a first aspect, a method for diagnosing a subtype of dementia in one or more patients is provided, the method comprising: obtaining brain imaging data related to the one or more patients; and classifying the one or more patients into a plurality of categories, including a first category of patients suffering from a first subtype of dementia and a second category of patients suffering from a second subtype of dementia, by providing the brain imaging data related to the one or more patients as input to a deep learning model, the deep learning model having been trained using brain imaging data from the plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal regions, or in addition to the temporal and parietal regions, wherein the first set of images is labeled as being associated with the first subtype of dementia, and the second set of images is labeled as being associated with the second subtype of dementia.

[0007] The inventors have found that metabolic brain imaging data from patients with AD and mixed AD and CVD surprisingly show significant differences in many brain regions that can be semi-quantitatively quantified and used to classify patients between these two subtypes using deep learning image classification models, although the coexistence and heterogeneity of pathology in dementia patients makes it difficult to distinguish typical AD from mixed pathology using existing methods.

[0008] The present invention also describes a method for analyzing brain imaging data from a patient, the method comprising: obtaining brain imaging data related to the patient; analyzing the data using a deep learning model that has been trained using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporal and parietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal areas instead of or in addition to the temporal and parietal areas, wherein the first set of images is labeled as being associated with a first subtype of dementia and the second set of images is labeled as being associated with a second subtype of dementia; and using the deep learning model, classifying one or more patients into a plurality of categories, including a first category of patients suffering from the first subtype of dementia and a second category of patients suffering from the second subtype of dementia.

[0009] Acquiring brain imaging data associated with one or more patients may include a processor receiving the brain imaging data associated with the one or more patients. The steps of analyzing and classifying may be performed by the processor.

[0010] The first and second sets of images and their associated labels together form the training dataset.

[0011] The deep learning model may provide as an output an indication of the probability that the patient from whom the image was acquired belongs to the first class and / or the second class.

[0012] Unless the context indicates otherwise, the methods described herein are computer-implemented. In fact, image analysis using deep learning models and the process of training deep learning models are complex, especially requiring large amounts of data to be analyzed through complex mathematics, which makes the methods described herein far beyond the capabilities of mental investigation.

[0013] In an embodiment, the subtype of dementia is a subtype of Alzheimer's disease (AD). The first subtype of dementia may be AD. The second subtype of dementia may be AD combined with cerebrovascular disease or vascular dementia. The second type of dementia may be cerebrovascular disease or vascular dementia. The second subtype of dementia may be mixed AD. The second subtype of dementia may be non-AD dementia.

[0014] Also described herein are methods and systems for diagnosing an AD subtype in one or more patients, the methods comprising: obtaining brain imaging data associated with the one or more patients; and classifying the one or more patients into a plurality of categories, including a first category of patients suffering from a first subtype of AD and a second category of patients suffering from a second subtype of AD, by providing the brain imaging data associated with the one or more patients as input to a deep learning model, the deep learning model having been trained using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal regions, or in addition to the temporal and parietal regions, wherein the first set of images is labeled as being associated with the first subtype of AD, and the second set of images is labeled as being associated with the second subtype of AD.

[0015] In an embodiment, the classification is between a first class of patients having a first subtype of dementia and a second class of patients having a second subtype of dementia. Thus, the classification can be a binary classification. For example, the classification can distinguish between patients having AD and patients having another subtype of dementia (e.g., mixed AD or AD and CVD).

[0016] In an embodiment, the deep learning model is a deep neural network classifier. In an embodiment, the deep learning model comprises a convolutional neural network (CNN). In an embodiment, the deep learning model comprises a model that has been pre-trained on unrelated image data. In an embodiment, the deep learning model comprises a CNN that has been pre-trained using a deep residual learning framework.

[0017] Convolutional neural networks have been shown to perform particularly well in image recognition tasks. Deep learning models may include models that have been pre-trained on large sets of image data for image recognition tasks (e.g., the available ImageNet database). These CNNs can be partially retrained on new data, for example by "freezing" (i.e., not retraining) lower-level layers (e.g., convolutional layers) that have been trained to identify lower-level features in images or only fine-tuning these layers, and only training or retraining higher-level layers (e.g., classification layers) to identify higher-level features that are particularly useful for the classification problem at hand. This partial retraining means that a limited amount of data can be used to quickly train deep CNNs, because only a subset of the parameters of the CNN need to be determined by training (in the case of freezing) and / or only the already optimized parameters need to be fine-tuned. This may be particularly advantageous when the amount of data available for training a specific classification task at hand is difficult and / or labor-intensive to obtain. Deep residual learning is a learning framework developed for image recognition to address a problem known as "descent" (observation that accuracy saturates and then rapidly decreases as the depth of the network increases). The deep learning model may include a pre-trained CNN that has been trained using deep residual learning (also known as ResNet). In an embodiment, the CNN is ResNet18. ResNet18 is a CNN that has been trained on more than one million images from the ImageNet database and in its native form (before retraining) can classify images into 1000 object categories, including, for example, keyboards, pencils, many animals, etc. The CNN is partially retrained to perform different image classification tasks (a process known as "transfer learning"), as described herein.

[0018] In an embodiment, the deep learning model includes all layers of a CNN that has been pre-trained for image recognition except for a classification layer, and a classification layer trained using the first and second sets of images and associated labels. In an embodiment, the classification layer includes a fully connected layer and a softmax layer.

[0019] In an embodiment, the images in the first and second sets show evidence of differential metabolic activity in any one or more or all of the following: right frontal cortex, left frontal cortex, right temporal cortex, left temporal cortex, right parietal cortex, left parietal cortex, left cerebellum, and right cerebellum. In an embodiment, the images in the second set of images show evidence of hypometabolism in regions of the brain that include one or more or all of the following: right frontal cortex, left frontal cortex, left cerebellum, and right cerebellum, instead of or in addition to one or more of the following: right temporal cortex, left temporal cortex, right parietal cortex, and left parietal cortex.

[0020] In embodiments, hypometabolism refers to a glucose uptake rate and / or blood flow and / or FDG-PET derived Standardized Uptake Value Ratio (SUVR) that is lower than expected for a control. A control may be a standard representative of a healthy patient.

[0021] In an embodiment, the brain imaging data is imaging data acquired using any functional brain imaging modality that provides information about the metabolic activity of the imaged brain region. In an embodiment, the information about the metabolic activity of the imaged brain region is acquired by detecting glucose uptake in the imaged brain region and / or blood flow to the imaged brain region.

[0022] In an embodiment, the brain imaging data is FDG-PET data or ASL data. In an embodiment, the brain imaging data is baseline functional brain imaging data.

[0023] In embodiments, analyzing the data using the deep learning model comprises analyzing a single cross-section of brain imaging data for each patient. In embodiments, the method comprises selecting a single cross-section of a set of brain imaging data for each patient. In embodiments, the single cross-section is a single axial cross-section at the level of the thalamus and / or wherein the single cross-section is a cross-section including at least a portion of the hippocampus and the entorhinal cortex.

[0024] In an embodiment, the brain imaging data used to train the deep learning model includes a single cross-section of a set of brain imaging data for each of a plurality of patients. In an embodiment, the method includes selecting a single cross-section of a set of brain imaging data for each of a plurality of patients. In an embodiment, the single cross-section is a single axial cross-section at the level of the thalamus. In an embodiment, the single cross-section is a cross-section including at least a portion of the hippocampus and the entorhinal cortex. A single cross-section refers to a single image, and the two terms are used interchangeably. The inventors have determined that a single cross-section of a set of brain imaging data can be used to accurately classify a patient as having AD or mixed AD. They further determined that this cross-section should preferably be an axial cross-section at the level of the thalamus.

[0025] In an embodiment, the method may be repeated for one or more additional cross-sections, and the classifications from each of the analyzed cross-sections may be combined to obtain a classification for the patient. Alternatively, the deep learning model may be configured to take a plurality of images as input and produce a classification of the plurality of images as output. For example, the deep learning model may include a plurality of instances of a deep learning model as described herein, each instance taking a single image as input, and a function that combines the outputs of these models to obtain a classification of the plurality of images.

[0026] In embodiments, the first set of images comprises one or more images for each of at least 30, at least 40, at least 50, up to 500, up to 200, or up to 100 patients. In embodiments, the second set of images comprises one or more images for each of at least 30, at least 40, at least 50, up to 500, up to 200, or up to 100 patients. In embodiments, the first set of images comprises one or more images for each of a number of patients that are within 10% or within 20% of the number of patients of the one or more images included in the second set of images. In embodiments, the first and second sets of images comprise one or more images for each of a plurality of patients, wherein the plurality of patients of the first and second sets of images are age-matched and gender-matched.

[0027] The inventors surprisingly discovered that a relatively small amount of training data having the characteristics described herein (i.e., specific patterns of hypometabolism) can be used to train a deep learning network to accurately distinguish subtypes of dementia.

[0028] Age and gender matched patient groups can refer to patient groups with approximately the same mean and / or distribution of age and gender. For example, the mean ages of the patient groups can be within 10% of each other, and / or the proportions of female and male patients can be within 10% of each other, and / or the proportions of patients in one or more age categories can be within 10% of each other.

[0029] In an embodiment, through expert review of the first and second sets of images, the first set of images has been labeled as being associated with a first subtype of dementia, and the second set of images has been labeled as being associated with a second subtype of dementia, optionally wherein each image in the first and second sets of images has been assigned the same label by at least two experts. Thus, if at least two independently acquired expert-derived labels are the same, then the images in the first set of images may be assigned a label associated with the first subtype of dementia. Similarly, if at least two independently acquired expert-derived labels are the same, then the images in the second set of images may be assigned a label associated with the second subtype of dementia. Images that are not assigned the same label by at least two experts may be excluded from the first and second sets of images.

[0030] In an embodiment, the brain imaging data used to train the deep learning model includes images from a plurality of patients and images obtained from the images from the plurality of patients by image enhancement. In an embodiment, the image enhancement includes creating a flipped version of one or more of the images and / or creating a randomly rotated version of one or more of the images. For example, a set of images (e.g., one or more or all of the original images) may be rotated by a randomly selected amount between predetermined boundaries, such as -30 to +30 degrees. As another example, a set of images (e.g., one or more or all of the original images) may be flipped horizontally.

[0031] In an embodiment, the deep learning model classifies the image into the first class and the second class with an accuracy of at least 90% or at least 95%. In an embodiment, the deep learning model classifies the image into the first class and the second class with a sensitivity of at least 90% or at least 94%. In an embodiment, the deep learning model classifies the image into the first class and the second class with a specificity of at least 90% or at least 95%. The accuracy, specificity, and / or sensitivity of the deep learning model classification can be measured by each of the following: performing cross-validation, such as 5 or 10-fold cross-validation, and quantifying the accuracy, specificity, and / or sensitivity of each split of the cross-validation, and optionally as an average or other summary measure of multiple splits.

[0032] In an embodiment, the method further comprises training the deep learning model using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in regions of the brain other than the temporal and parietal regions instead of or in addition to the temporal and parietal regions, wherein the first set of images are labeled as being associated with a first subtype of dementia and the second set of images are labeled as being associated with a second subtype of dementia.

[0033] According to a second aspect, a method for providing a tool for diagnosing a subtype of dementia in one or more patients is provided, the method comprising: obtaining training image data, the training image data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal areas instead of or in addition to the temporal and parietal areas, wherein the first set of images are labeled as being associated with a first subtype of dementia and the second set of images are labeled as being associated with a second subtype of dementia; and training a deep learning model to classify patients into a plurality of categories, including a first category of patients suffering from the first subtype of dementia and a second category of patients suffering from the second subtype of dementia.

[0034] The method of aspects of the present invention may have any of the features described with respect to the first aspect. The method of aspects of the present invention is preferably computer-implemented. As described above, at least the step of training the deep learning model is computer-implemented in any practical application. Therefore, the steps of the method may include a processor executing instructions to perform the steps. For example, obtaining training image data may include a processor executing instructions to obtain training image data from a data source (e.g., a database, computer memory, etc.). Similarly, training a deep learning model may include a processor executing instructions to train a deep learning model.

[0035] Training the deep learning model may include at least partially retraining a pre-trained deep learning model. Partially retraining the deep learning model may include fixing parameters of one or more of the lower layers of the model, and determining parameters of the remaining (higher level) layers of the model. Partially retraining the deep learning model may include fine-tuning weights of a plurality of layers of the deep learning model and training weights of one or more additional layers. The one or more additional layers may be classification layers. The classification layers may include fully connected layers and softmax layers.

[0036] According to any aspect described herein, acquiring brain imaging data may include receiving brain imaging data from a computing device, an imaging data acquisition device, a data storage device, or a user interface. The method of any aspect may include acquiring brain imaging data from a patient. In some cases, these methods include administering an imaging tracer to a patient and acquiring brain imaging data from the patient. This step may not be computer-implemented and may precede any computer-implemented steps performed on the acquired data. Alternatively, all steps of the method may be computer-implemented and include receiving previously acquired brain imaging data.

[0037] The method of any aspect may include providing the results of the classification, the trained deep learning model, and / or any information derived therefrom to the user, for example, through a user interface. The data storage may be a public or private database. The results of the classification may include one or more of the following: the probability of belonging to the first class and / or the second class obtained using the deep learning model, the classification labels of one or more images, the classification labels of one or more patients, the trained deep learning model, the parameter values ​​of the trained deep learning model (e.g., architecture and weights). The information derived from the results of the classification may include one or more of the following: a prognostic indication derived from the classification obtained using the deep learning model, a treatment indication derived from the classification obtained using the deep learning model, an indication of suitability for participation in a clinical trial derived from the classification obtained using the deep learning model, etc.

[0038] According to a third aspect, a method for selecting individuals who have been diagnosed with AD or who may have AD to participate in a clinical trial is provided, the method comprising: classifying the individuals into a plurality of categories, including a first category of patients with a first subtype of dementia and a second category of patients with a second subtype of dementia, by providing brain imaging data related to the individuals as input to a deep learning model, the deep learning model having been trained using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporal and parietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal areas, or in addition to the temporal and parietal areas, wherein the first set of images is labeled as being associated with the first subtype of dementia and the second set of images is labeled as being associated with the second subtype of dementia; and selecting or excluding individuals from participating in the clinical trial depending on whether the individuals are classified as having the first subtype of dementia or the second subtype of dementia.

[0039] The method of aspects of the present invention may have any of the features described with respect to the first aspect. In particular, if the individual is classified as the first category, the method may include selecting the individual to participate in a clinical trial. If the individual is classified as the second category, the method may include excluding the individual from participating in the clinical trial. The clinical trial may be a trial for the treatment of AD. The method of aspects of the present invention may include any combination of some, all, or none of the above-mentioned preferred and optionally selected features.

[0040] According to a fourth aspect, a method for providing a prognosis for an individual who has been diagnosed with AD or is likely to have AD is provided, the method comprising: classifying the individual into a plurality of categories, including a first category of patients with a first subtype of dementia and a second category of patients with a second subtype of dementia, by providing brain imaging data related to the individual as input to a deep learning model, the deep learning model having been trained using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporal and parietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal areas, or in addition to the temporal and parietal areas, wherein the first set of images is labeled as being associated with the first subtype of dementia and the second set of images is labeled as being associated with the second subtype of dementia; and determining a prognosis for the individual based on the individual being classified as having the first subtype of dementia or the second subtype of dementia, where determining the prognosis comprises determining that if the individual is classified as the second category, the individual is likely to have a faster rate of cognitive decline than if the individual is classified as the first category. The method of aspects of the present invention may include any combination of some, all or none of the above-mentioned preferred and optional features.

[0041] According to a fifth aspect, a system for diagnosing a subtype of dementia and / or a tool for diagnosing a subtype of dementia is provided, the system comprising: one or more processors and a computer-readable memory storing instructions, the instructions causing the processor to execute a method of any embodiment of any of the aforementioned aspects, in particular including any embodiment of the first aspect and / or the second aspect. The system may further include data acquisition devices configured to acquire brain imaging data associated with one or more patients. In some embodiments, the system may include, for example, one or more computers, servers, or cloud-based devices.

[0042] According to a sixth aspect, there is provided a non-temporary computer-readable storage medium containing machine-executable instructions which, when executed on a processor, cause the processor to perform a method of any implementation of aspects one to four, including any one or any combination of the optional features listed therein, provided that they are compatible.

[0043] According to the seventh aspect, a computer program comprising executable codes is provided, which, when run on a computer, causes the computer to execute the method of any implementation of any of the preceding aspects, specifically including any implementation of the first to fourth aspects, including any one or any combination of the optional features listed with reference thereto, as long as they are compatible.

[0044] The present invention includes any combination of described aspects and preferred features unless such a combination is expressly impermissible or explicitly avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Embodiments and experiments illustrating the principles of the present invention will now be discussed with reference to the accompanying drawings, in which:

[0046] Figure 1 Examples of each of the following are shown: an FDG-PET image showing evidence of temporoparietal hypometabolism, also referred to herein as the "classic AD pattern" of hypometabolism (A), and an image showing evidence of hypometabolism in areas of the brain other than the temporal and parietal regions, also referred to herein as the "mixed pattern" of hypometabolism (B), specifically patchy hypometabolism in other vascular areas of the brain (e.g., frontal and cerebellar regions) as well as temporoparietal hypometabolism. A = anterior, P = posterior, L = left, R = right, H = superior, F = inferior.

[0047] Figure 2 The architecture of the deep learning model according to the embodiment of the method of the present disclosure is schematically shown. The figure particularly shows the modified ResNet-18 architecture used in the examples of the present disclosure.

[0048] 3 shows an occlusion sensitivity map of an example deep learning model in the present disclosure for classification of a patient of a first subtype showing a typical AD hypometabolism pattern (A) or a patient of a second subtype showing a hypometabolism mixed pattern (B).

[0049] Figure 4 Flowchart of a method for diagnosing Alzheimer's disease subtypes and / or for providing a tool for diagnosing AD subtypes.

[0050] Figure 5 FIG. 4 is a schematic diagram of a system for executing the method of FIG. 4 . DETAILED DESCRIPTION OF THE INVENTION

[0052] Aspects and embodiments of the present invention will now be discussed with reference to the accompanying drawings. Other aspects and embodiments will be apparent to those skilled in the art. All documents mentioned herein are incorporated herein by reference.

[0053] The inventors analyzed brain images (specifically FDG-PET scans) of patients with two different types of AD (AD and mixed AD and CVD) and showed for the first time that the two subtypes can be identified using semi-quantitative image analysis, are clinically distinct, and can be distinguished with very high accuracy using automated image analysis using a deep neural network model trained by transfer learning.

[0054] Approach to diagnosing dementia subtypes

[0055] Figure 4A flow chart showing a method for diagnosing a subtype of dementia in a target patient and a method for providing a tool for diagnosing a subtype of dementia is shown. The method may be performed on one or more computing devices. In step 400, training brain imaging data is obtained from a plurality of patients. The training brain imaging data may be baseline FDG-PET scan data. In step 410, which is optional, a single cross-section (image) is selected for each of the plurality of patients. In step 420, if the single cross-section shows evidence of temporoparietal hypometabolism, a first label is assigned to each of the single cross-sections, and if the single cross-section shows evidence of hypometabolism in a region of the brain other than the temporal and parietal regions instead of or in addition to the temporal and parietal regions, a second label is assigned. Step 420 may be performed by expert visual analysis, preferably including obtaining consistent labels from at least two different experts in step 425. In step 430, which is optional, a subset of the training brain imaging data may be selected, for example to include a similar number of images assigned with the first and second labels, and / or to include age- and gender-matched images assigned with the first and second labels. At step 435, which is optional, the training brain imaging data or a subset thereof may be image enhanced, for example, by flipping and / or rotating one or more of the images. At step 440, the training brain image data (optionally selected and / or enhanced) and labels are used to train a deep learning model. Training the deep learning model may include at least partially retraining a pre-trained deep learning model. The pre-trained deep learning model may have been previously trained to perform an unrelated image recognition task. Thus, step 440 may include, at step 445, obtaining a pre-trained deep learning model, for example, from a memory, a data storage, a user interface, or a computing device. At step 450, which is optional, the performance of the deep learning model may be evaluated, for example, by determining its accuracy, sensitivity, and / or specificity in distinguishing between images assigned a first label and images assigned a second label. This may be performed using cross-validation, such as a 5-fold cross-validation. Thus, for each iteration of the training and validation process, the training image data may be divided between the training and validation data sets. At step 460, which is optional, the trained deep learning model may be provided to a user. At step 470, brain imaging data associated with one or more patients is obtained. At step 480, optionally, a single cross-section (image) is selected for each of the one or more patients. At step 490, a deep learning model, such as obtained through steps 400-460, is used to classify the one or more patients between a plurality of categories, the categories including a first class of patients with a first subtype of dementia and a second class of patients with a second subtype of dementia. At step 500, optionally, the classification results or information derived therefrom are provided to a user. The information derived from the classification results may include one or more of diagnosis, prognosis, treatment, clinical trial selection, as will be further described below.The method described herein may include any of steps 400 to 460 and / or any of steps 470 to 500. The selection of a single image in steps 410 and / or 480 may be performed manually or automatically. For example, a predetermined cross section of any brain image dataset may be selected (e.g., a single axial cross section including a predetermined one or more regions of interest such as the thalamus and / or hippocampus and / or entorhinal cortex may be selected).

[0056] The terms "subject" and "patient" are used interchangeably in this disclosure.

[0057] Use of analysis results

[0058] The results of this analysis can be used to diagnose whether the patient has AD or mixed AD, such as AD and cerebrovascular disease (CVD). The results of this analysis can be used to select patients for participation in clinical trials. For example, a clinical trial can be designed to exclude patients with mixed AD and / or only include patients with "typical" (i.e., non-mixed) AD. As another example, a clinical trial can be designed to exclude patients with "typical" AD and / or only include patients with "mixed" AD. Also described herein is a method for selecting individuals who have been diagnosed with AD or may have AD to participate in a clinical trial, the method comprising: analyzing one or more brain images from a patient using the method described herein, and selecting or excluding individuals from participating in the clinical trial depending on whether the patient is classified as having a first dementia (or AD) subtype or a second dementia (or AD) subtype.

[0059] The results of this analysis can be used to predict the level of cognitive impairment in patients who have been diagnosed with AD or may have AD, where patients classified as having mixed AD and CVD may have more severe cognitive impairment than patients with AD. The results of this analysis can be used to provide a prognosis for patients who have been diagnosed with AD or may have AD. In fact, patients with mixed AD coexisting with AD and CVD show a faster rate of cognitive decline than patients with AD alone (see, for example, Zekry et al., Acta Neuropathol 2002; 103: 481-7 and Kapasi et al., Acta Neuropathol 2017; 134: 171-86). Therefore, the method of the present invention can be used to predict whether a patient may have a faster rate of cognitive decline than expected in AD patients. In addition, if a treatment that can treat vascular pathology becomes available, the method can be used to help select patients suitable for treatment. For example, Rogriguez et al. (Brain Res 1588 (2014): 144-149) reported that methylene blue can reduce the extent of hypoxemic damage to brain tissue caused by occlusion of the carotid artery supplying the brain. Therefore, the present invention also describes a method for providing a prognosis for an individual who has been diagnosed with AD or may have AD, the method comprising: analyzing one or more brain images from a patient using the method described herein, and identifying the prognosis of the individual based on whether the patient is classified as suffering from a first subtype of dementia or a second subtype of dementia. For example, a patient classified as a patient category with a second subtype of AD may be associated with a poor prognosis compared to a patient classified as a patient category with a first subtype of AD. A poor prognosis may refer to a faster than expected rate of cognitive decline in an AD patient. The expected rate of cognitive decline in an AD patient may be the average rate of cognitive decline observed in an AD patient group. An AD patient group may be an age- and sex-matched patient who has been diagnosed with AD. An AD patient group may be a patient diagnosed with AD in the absence of CVD. A patient group may be a patient diagnosed with AD by analyzing an FDG-PET image. A patient group may be a patient with only temporoparietal hypometabolism in an FDG-PET image.

[0060] The results of such an analysis can be used to treat or identify patients with dementia. For example, a patient diagnosed with "classic" dementia may be treated differently from a patient diagnosed with mixed dementia, which includes AD and vascular dementia / CVD. For example, a patient with mixed dementia may be treated with compounds that treat high blood pressure, lower cholesterol, and / or prevent blood clots instead of or in addition to compounds used to treat AD. As another example, a patient with mixed dementia may be advised to make lifestyle changes, such as changes in diet, activity patterns, alcohol consumption, or smoking, instead of or in addition to treatment for AD. As described above, a patient with mixed dementia may be recommended or selected for treatment with a therapy that treats vascular pathology (e.g., methylene blue). As another example, a patient with dementia may be treated with a compound that treats AD only after vascular dementia has been ruled out and / or AD has been diagnosed. Thus, also described herein is a method for identifying a therapy for an individual who has been diagnosed with dementia or who may have dementia, the method comprising: analyzing one or more brain images from a patient using the methods described herein, and identifying the individual for treatment with a first therapy or a second therapy depending on whether the patient is classified as having a first dementia (or AD) subtype or a second dementia (or AD) subtype. The first therapy may be a therapy for treating AD. The second therapy may be a therapy for treating vascular dementia. The second therapy may include a therapy for treating vascular dementia and a therapy for treating AD. Also described herein is a method for selecting an individual with dementia to be treated with a vascular dementia therapy, the method comprising analyzing one or more brain images from a patient using the methods described herein, and if the patient is classified as having a second dementia subtype, selecting the individual to be treated with a vascular dementia therapy. Also described herein is a method for selecting an individual with dementia to be treated with an AD therapy, the method comprising analyzing one or more brain images from a patient using the methods described herein, and if the patient is classified as having a first dementia subtype, selecting the individual to be treated with an AD therapy. Also described herein is a method of treating an individual having dementia, the method comprising analyzing one or more brain images from the individual using the methods described herein, and administering to the individual a therapeutically effective dose of a therapy for treating vascular dementia if the individual has been classified as a second type of patient having a second subtype of dementia, and / or administering to the individual a therapeutically effective dose of a therapy for treating AD if the individual has been classified as a first type of patient having a first subtype of dementia.

[0061] As used herein, "treat," "treat," and "treatment" refer to the reduction, alleviation, or elimination of one or more symptoms of the disease being treated, relative to pre-treatment symptoms.

[0062] system

[0063] Figure 5An embodiment of a system for diagnosing AD subtypes and / or classifying individuals as suffering from AD or mixed dementia and / or analyzing brain images and / or providing tools for diagnosing AD subtypes according to the present disclosure is shown. The system includes a computing device 1, which includes a processor 101 and a computer-readable memory 102. In the illustrated embodiment, the computing device 1 also includes a user interface 103, which is shown as a screen but may include any other means of conveying information to the user, such as through sound or visual signals. The computing device 1 is communicatively connected to a data acquisition device 3 (also referred to as a "brain image data acquisition device"), such as a PET machine or an MRI machine or a computing device associated therewith, for example, via a network, and / or to one or more databases 2 storing brain imaging data. The one or more databases 2 may further store one or more of the following: one or more deep learning algorithms, training data, parameters (such as parameters of a deep learning model for diagnosing AD subtypes), clinical and / or sample-related information, etc. The computing device may be a smartphone, a tablet computer, a personal computer, or other computing device. As described herein, the computing device is configured to implement methods for diagnosing AD subtypes, analyzing brain images, and / or classifying individuals as suffering from AD or mixed dementia. In alternative embodiments, the computing device 1 is configured to communicate with a remote computing device (not shown), which itself is configured to implement the methods as described herein. In such cases, the remote computing device may also be configured to send the results of the method to the computing device. The communication between the computing device 1 and the remote computing device may be performed via a wired or wireless connection, and may occur on a local or public network 6, such as the public Internet. The data acquisition device 3 may be wired to the computing device 1, or may be able to communicate via a wireless connection (e.g., via a WiFi connection) and / or on the public Internet, as shown. The connection between the computing device 1 and the data acquisition device 3 may be direct or indirect (e.g., via a remote computer). The data acquisition device 3 is configured to collect brain imaging data from the patient. Any imaging scheme suitable for obtaining information about metabolic activity in the patient's brain (e.g., PET and MRI) may be used in the context of the present invention. The data acquisition device preferably comprises a PET scanner, preferably configured to collect FDG-PET images. The data acquisition device may comprise arterial spin labelling (ASL).

[0064] The following content is presented by way of example and should not be construed as limiting the scope of the claims.

[0065] Examples

[0066] As described above, the use of visual interpretation of FDG-PET images and the use of semi-quantitative methods to explore and discover the limitations of traditional methods for detecting and monitoring AD from such images can be minimized. Recently, automated techniques, such as artificial intelligence (AI), have been proposed as a potential alternative (Jo et al., Front.AgingNeurosci.2019; 11). Deep learning is a machine learning technique that can be trained directly using images, text or sound to learn the classification format of a given input. The inventors assume that these techniques can be applied using expert classification of AD subtypes as a gold standard, and classification algorithms can be trained to distinguish AD from mixed dementia. For example, deep learning algorithms such as convolutional neural networks (CNNs) have been used to analyze medical images, which obtain input images, analyze them according to the trained algorithms, and classify them into specific categories (Yadav et al., J Big Data 2019; 6 (1): 113). However, training a new CNN is challenging and computationally demanding. Model validation makes this method impractical for medium-sized data sets with sample sizes of hundreds, because it requires large data sets with sample sizes of thousands to obtain acceptable accuracy. To address this problem, image classification algorithms can be trained using large, general-purpose image datasets and then “fine-tuned” using appropriate medical image data. This approach is called transfer learning and is an improvement on CNNs that can be applied to smaller datasets.

[0067] The introduction of advanced methods to classify AD into subtypes will improve the reliability of diagnosis and may help clinical decision-making in the management of AD patients. Here, the inventors used visual classification of FDG-PET images to divide participants with AD into participants with typical FDG-PET low metabolism AD patterns (referred to as "typical") and participants with typical FDG-PET low metabolism plus CVD typical FDG-PET low metabolism AD patterns (referred to as "mixed"). Then, they evaluated the differences between these groups by SUVR (standardized uptake value ratio) analysis based on the region of interest. In addition, they constructed a classification model based on the transfer learning of the Residual Network-18 (ResNet-18) architecture (He et al. 2016IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2016: 770-778) to classify images, treating visual classification as the gold standard. Using data from two large clinical trials of people with good features that meet the clinical criteria for AD, the inventors aimed to use conventional visual classification to classify people as having typical or mixed AD, and then compared these classifications with the classification of transfer learning using convolutional neural networks.

[0068] method

[0069] Selection and recruitment of participants. The present invention uses Leuco-Methylthioninium as a novel tau aggregation inhibitor Baseline FDG-PET data were collected as part of two large-scale Phase III clinical trials of 1690 participants who met the study criteria for mild to moderate AD and were younger than 90 years of age. 794 participants were included in the current study, and 896 participants were excluded due to lack of FDG-PET imaging data or incomplete images. Details of the clinical trial approval and methods have been described in previous publications (see Gauthier et al. The Lancet. 2016; 388(10062): 2873-2884; Wilcock et al., J Alzheimers Dis. 2018; 61(1): 435-457; Schelter et al., J Alzheimers Dis. 2019; 72(3): 931-946). Written informed consent was obtained from each participant before participation in studies TRx-237-015 (clintrials.gov NCT01689246) and TRx-237-005 (clintrials.gov NCT01689233). For patients who lacked decision-making capacity, their consent was provided by a legal representative.

[0070] FDG PET imaging protocol. Brain images were acquired by positron emission tomography computerized tomography (PET CT) using high-resolution PET devices, such as the Siemens High Resolution Research Tomograph (HRRT) system or the GE Medical Systems Discovery STE, which use a transmission source or low-dose CT for attenuation correction. Prior to enrolling participants, the PET sites had to undergo strict quality standards set and checked by Molecular NeuroImaging (mnimaging.com). A standard dose of 5mCi / 185MBq (±10%) FDG was injected intravenously into the antecubital region over a period of 1 minute via an indwelling catheter. PET images were acquired 30 minutes (±5 minutes) after administration of FDG.

[0071] Image reconstruction and processing. Images were reconstructed using a 128x128 image matrix with a slice thickness of 4 mm. Images were scaled with a field of view (FOV) of 350 mm and a Gaussian full width at half maximum (FWHM) smoothing of 5.0 mm. Standard low-dose non-diagnostic CT acquisitions of the head were acquired for attenuation correction. Images were evaluated for artifacts, patient motion, excessive noise, low counts, and patient positioning.

[0072] Visual Classification of FDG-PET Images. FDG-PET images were classified using visual inspection into images with a typical AD pattern with temporoparietal hypometabolism (see, e.g., Marcus et al. Clin Nucl Med 2014;39:e413-26; Milke et al., Eur J Nucl Med 1994;21:1052-60; Kerrouche et al., J Cereb Blood Flow Metab 2006;26:1213-21; Garibotto et al., Neurobiol Aging 2017;52:183-95) and images with a mixed AD and CVD pattern (mixed pattern) with temporoparietal hypometabolism and defects in one or more vascular territories (i.e., known blood supply areas of the brain, such as the middle cerebral artery). Scans with typical AD FDG-PET features had reduced glucose uptake limited to the temporoparietal regions (see Figure 1A) In addition to typical temporoparietal hypometabolism, scans with mixed AD / CVD features have reduced FDG uptake in specific vascular territories (e.g., middle cerebral artery and / or patchy uptake) (see Figure 1 B). PMOD Alzheimer's disease discrimination analysis tool (PALZ, Haense et al., Journal of Nuclear Medicine. 2008; 49 (Suppl. 1): 34P-34P) is used for visualization of FDG-PET images. Classification is based on visual inspection of scanned images displayed on three planes using a standard color scale representing FDG absorption. In order to determine the level of variability between scorers, a subset of the data is classified by a second trained observer, and Cohen's Kappa is calculated. The degree of agreement between the scorers is moderate. The inter-scorer reliability measured using Cohen's kappa is 0.55, indicating that the degree of agreement is acceptable. In the case of inconsistency, the images are reviewed together and discussed until a consensus is reached. The final consensus classification obtained by the two scorers is used as the gold standard for training the machine learning model (see below). Figure 1 Examples of FDG pet images of the brains of participants classified as each of: A) a typical AD pattern of glucose metabolism with areas of temporoparietal hypometabolism 10 and B) a mixed pattern with patchy hypometabolism in other vascular areas of the brain (e.g., frontal and cerebellar regions 12) as well as areas of temporoparietal hypometabolism 10.

[0073] Analysis of visually classified FDG-PET images. Quantification of FDG-PET metabolism was achieved by determining the standardized uptake value ratios (SUVRs) of different brain regions, namely the left and right frontal cortex, left and right temporal cortex, left and right parietal cortex, left and right occipital cortex, and left and right cerebellar cortex, with their intensity normalized to the pons (Nugent et al., Scientific Reports. 2020; 10(1):9261).

[0074] Description of the transfer learning method. For the purpose of transfer learning, 50 age- and sex-matched participants were selected from the AD and mixed groups of visual classification using covariate adaptive randomization. It should be noted that the learning dataset does not necessarily contain the same number of participants in the two groups, or participants of age and sex matching. However, the use of age- and sex-matched participants may increase the reliability of the resulting classifier in terms of its performance in the validation test (i.e., how its performance is generalized to datasets other than the training dataset). Using SPM12 implemented in MATLAB R2020a (available at www.fil.ion.ucl.ac.uk / spm / software / spm12), the participants' FDG-PET images were standardized to PET templates in the standard Montreal Neurological Institute (MNI) space (the PET template included in SPM12). A single axial section at the level of the thalamus was used as input for ResNet-18 transfer learning. It should be noted that any image, in particular any axial section, can be used, including areas around the hippocampus and entorhinal cortex. Image enhancement is performed by creating a flipped version of each axial section (horizontally flipped) and then randomly rotating the section by -30 to 30 degrees. Image enhancement advantageously reduces the risk of overfitting the network to the training images. Twenty data sets (i.e., single images from 20 patients) were retained for testing, and the remaining 80 data sets were divided into training and validation data sets in a ratio of 80% / 20%. The final classification layer of ResNet-18 was replaced by a new fully connected layer, which contained 2 classes to represent the mixed group and the AD group. ResNEt-18 is an 18-layer deep convolutional neural network (CNN). The version used was pre-trained on more than one million images from the ImageNet database (www.image-net.org) and implemented in MatLab. Figure 2 Figure 1 shows the data flow stages of the machine learning algorithm used in this study. In particular, Figure 2The basic architecture of ResNet-18 is shown, which shows the different layers of CNN. The figure shows skip connections (skipping some layers in the neural network and feeding the output of one layer as input to the next layer) in the form of dotted lines and numbers describing the output size. Avg pool is a global average pooling layer, which reduces the spatial size, computational complexity and parameter number of representation, and FC is a 1000-way fully connected layer, which is a feedforward neural network with 1000 FC sublayers and fully connected with all previous sublayers. The last layer is a softmax layer, which is the last activation function of the neural network, and is used for the output of the standard network to provide an output between 0 and 1. Five times of cross validation are adopted and the average accuracy of five times is reported. The trainNetwork function in MatLab is used to train the network, thus using training images to fine-tune the weight of the pre-trained network, rather than using the weight of random initialization to "start from scratch" training network. Although the size of the available training data set is relatively moderate, this advantageously leads to fast training and very good performance.

[0075] Statistical analysis. All statistical analyses were performed using SPSS Version-26. Descriptive statistics are provided, and means were compared where appropriate. Student's t-test was used to test mean differences for normally distributed continuous data, and the chi-square test was used to detect differences in binary data. General linear models were used to further explore relationships between variables where appropriate. P values ​​< .05 were considered significant, and the convention for indicating significance levels is * < .05, ** < .01, *** < .001. In addition, sensitivity, specificity, and accuracy between true and predicted subtypes of AD were calculated.

[0076] result

[0077] Table 1 shows the demographic and clinical characteristics of the participants in this study. The average age of the participants was 70.56 years old, and more women participated (55.16% vs. 44.83%). Of the total 794 (438 female) participants, 533 (284 females) were classified as typical AD, and 261 (154 females) were classified as mixed (Table 1). In addition, 100 age- and gender-matched participants (50 each of typical AD and mixed) were selected for transfer learning. Compared with typical AD participants, participants classified as having mixed hypometabolism were younger and had more severe cognitive impairment.

[0078]

[0079] Table 1: Comparison of demographic and clinical characteristics with visual FDG-PET AD classification.

[0080] Table 2 shows the differences in mean SUVRs for different regions of interest (ROIs) compared to typical AD and mixed subtypes. The ROIs used were permanent regions in the Montreal Neurological Institute (MNI) standard space, created in standard space and copied to a standardized version of the patient data. Significant differences were found when comparing the mean SUVRs of AD and mixed type subtypes in the left and right frontal, right temporal, and left and right parietal cortices. However, no significant differences were found in the left temporal cortex, left and right occipital cortices, and left and right cerebellum.

[0081]

[0082] Table 2. Comparison of regional SUVR between typical AD and mixed type. The significance level was adjusted according to the Bonferroni correction and was considered significant if *P < 0.005, **P < 0.001 and ns-not significant.

[0083] Furthermore, AD subtypes were compared with different regions of interest (ROIs), population weighted age and sex were controlled by using these in the comparisons in the statistical analysis (Table 3) and SUVR was observed to be lower in those with mixed patterns in all ROIs except for the left and right occipital cortices. SUVR in the right frontal cortex was statistically significant when associated with dementia subtype.

[0084]

[0085] Table 3. Comparison of regional uptake of typical AD and mixed subtypes, controlling for age and sex. Effect = β, Age = population-weighted age, AD subtype = AD type and mixed type, considered significant if P < 0.05.

[0086] The sensitivity, specificity, and accuracy of the ResNet-18 based classification model trained using transfer learning were found to be 94.73%, 95.23%, and 95% for a randomly selected cross validation cycle, respectively. The average accuracy after 5-fold cross validation was found to be 97.5%.

[0087] Figure 3A and 3B Occlusion sensitivity maps for six participants from the typical AD (A) and mixed type (B) groups are shown. The occlusion map shows which regions 30 of the image contribute positively to the group classification. Occlusion measures the network's sensitivity to occlusion of different regions of the image (the probability score of a particular class decreases) by replacing small regions in the image with occlusion masks (e.g., gray squares). Figure 3A and 3B It is shown that the most informative region for classifying patients between AD and mixed forms is Figure 3A The highlighted area is the area that is a strong indicator of a hypometabolic AD pattern.

[0088] discuss

[0089] This study is the first of its kind, not only to subclassify AD patients based on FDG-PET images, but also to introduce an automatic classification method that uses FDG-PET images analyzed by machine learning algorithms trained using traditional visual analysis data, especially using advanced transfer learning techniques to classify using FDG-PET images. The inventors visually classified people with AD into typical AD and mixed subtypes based on the low metabolic patterns seen in FDG-PET images, and regarded this classification as the gold standard. Based on the visual analysis of FDG-PET, the two subtypes of Alzheimer's disease differ significantly not only in visually determined glucose uptake patterns, but also in clinical and demographic characteristics, as people with the mixed subtype are younger and more cognitively impaired. To explore the results of the visual analysis, a semi-quantitative analysis was performed and SUVR was calculated in standard ROIs. Regions corresponding to metabolic defects in the visual analysis had lower SUVRs. It is noteworthy that in the current study, typical AD patients had higher glucose uptake compared to mixed patients in most ROIs, except for the occipital cortex, where no significant differences were found. Compared with typical AD patients, mixed patients have lower uptake rates, indicating that most areas of mixed patients have poor blood supply and are consistent with a general decrease in brain parenchymal metabolism. This is consistent with recent studies showing that people with both AD and cerebrovascular disease have more severe disease than those with AD alone (De Reuck J. Neurol Res Int. 2019; 2019: 7247325). In order to establish the utility of automatic FDG-PET classification, an advanced machine learning technique, namely transfer learning, is introduced. Using machine learning, especially transfer learning, to classify FDG-PET images of people with AD subtypes is novel. The ResNet-18 convolutional network with transfer learning was selected because it can provide accurate model construction in a very short period of time (Rawat and Wang, Neural Comput. 2017; 29 (9): 2352-2449).It should be noted that any deep neural network architecture suitable for image classification can be used, such as squeezenet (Iandola et al., arxiv.org / abs / 1602.07360), googlenet, inceptionv3 (Szegedy et al., Proc of IEEE conf comp vis pat recog, pp. 1-9, 2015), densenet201 (Huang et al., CVPR vol. 1, no. 2, p / 3, 2017), resnet-50 or -101 (He et al., Proc IEEE conf comp vis pat rec, pp. 770-778, 2016), efficientnetb0 (Migxing Tan and Quoc, Arxiv:1905.1194, 2019), alexnet (Krizhevsky et al., Adv neur info proc sys, 2012), vgg16 (Simonyan and Ziserman, arxiv:1409.1556, 2014), etc. The ResNet-18 convolutional network with transfer learning is able to distinguish the two subtypes, and the areas that contribute most to typical AD are indicated in red on the occlusion sensitivity map (. Figure 3A Reference number 30 in ). Although the machine learning model was trained on the whole image and did not receive any information about the region of interest (ROI) as input, it was able to pick out regions that could provide information about the difference between AD and mixed AD. After 5-fold cross-validation, the average accuracy of distinguishing the two subtypes was found to be 97.5%. The prediction accuracy of the current study is much higher than some previous studies based on deep learning technology (Korolev et al., arXiv:170106643[cs]. Published online on January 23, 2017; Lu et al., Sci Rep.2018;8), although previous studies have attempted to distinguish people with AD or mild cognitive impairment (MCI) from healthy controls (Table 4). The classification accuracy shown here for distinguishing AD patients from mixed dementia patients is similar to the best accuracy for distinguishing AD patients or cognitively impaired patients from healthy controls. The identification of AD subtypes is a more difficult problem than the classification of AD and control or MCI and control, or even AD and MCI. It is therefore surprising that the method of the present invention can achieve such a high accuracy in this much more difficult classification task.

[0090] The overall accuracy was higher in studies differentiating between AD and controls compared to studies differentiating between MCI and controls. This means that most models were not as accurate in differentiating the subtle differences between images of groups with some form of memory impairment (i.e. AD and MCI) as they were in cognitively healthy images. The current classifier based on ResNet-18 and transfer learning was able to distinguish between the classic and mixed form subtypes of AD with 97.5% accuracy, despite classifying two subgroups of the same disease.

[0091]

[0092] Table 4. Overview of recent AD research based on deep learning techniques. MRI = magnetic resonance imaging, PET = positron emission tomography, CNN = convolutional neural network, DNN = deep neural network, MMDNN = multi-scale multi-model deep neural network Acc = accuracy, SPE = specificity and SEN = sensitivity, AD = Alzheimer's disease, MCI = mild cognitive impairment, SVM = support vector machine, CSF = cerebrospinal fluid (measures such as 42 residues long amyloid-β isoform or total tau), APOE = apolipoprotein E.

[0093] This study introduced deep learning for the first time to classify FDG-PET images of AD patients into two clinically important subtypes, "typical AD" and "mixed". The overlapping mixed pathologies of AD patients pose challenges to clinicians' diagnosis and patient management. The introduction of automatic transfer learning technology to classify patients into typical AD and mixed subtypes not only helps to accurately distinguish between pure subsets of typical AD patients and mixed patients, but also provides the potential for more accurate, repeatable and faster diagnosis.

[0094] in conclusion

[0095] This study is the first to differentiate between two radiological subtypes of AD through visual analysis of FDG-PET images, and then demonstrates that the two subtypes are distinguishable from each other on a semiquantitative analysis and are clinically distinct. In addition, transfer learning, a type of machine learning, has been used to predict the two subtypes with high accuracy, sensitivity, and specificity. In clinical settings, most images are analyzed by experts through visual analysis, which is expensive and time-consuming, and has well-known intra- and inter-observer bias. Machine learning techniques like transfer learning can overcome these shortcomings. On a pragmatic level, they can also meet the growing demand given the common challenges of an aging population and a global shortage of radiologists. The new application of transfer learning, which utilized the pre-trained network ResNet-18 in this study, may greatly improve the efficiency and accuracy of FDG-PET images in differentiating between patients with typical AD and those with coexisting cerebral small vessel disease. This application of AI is not only beneficial for accurate diagnosis and prognosis of individual patients, but is also important for deciding the right patients to recruit into future clinical trials.

[0096] In addition to the described structural components and user interactions, the systems and methods of the above embodiments can also be implemented in a computer system (particularly computer hardware or computer software).

[0097] The term "computer system" includes hardware, software, and data storage devices for embodying the system or performing the method according to the above-described embodiments. For example, the computer system may include one or more processing units, such as a central processing unit (CPU) and / or a graphics processing unit (GPU), an input device, an output device, and a data storage device. Preferably, the computer system has a monitor to provide a visual output display. The data storage device may include RAM, a disk drive, or other computer-readable media. The computer system may include a plurality of computing devices connected by a network and capable of communicating with each other via the network. It is clearly contemplated that the computer system may consist of or include a cloud computer.

[0098] The methods of the above-described embodiments may be provided as a computer program or as a computer program product or a computer-readable medium carrying a computer program, wherein the computer program is configured to perform the above-described methods when executed on a computer.

[0099] The term "computer-readable media" includes, but is not limited to, any non-transitory media or media that can be directly read and accessed by a computer or computer system. The media may include, but is not limited to, magnetic storage media such as floppy disks, hard disk storage media, and tapes; optical storage media such as optical disks or CD-ROMs; electrical storage media such as memory, including RAM, ROM, and flash memory; and hybrids and combinations of the above, such as magnetic / optical storage media.

[0100] The features disclosed in the preceding description or in the following claims or in the accompanying drawings and expressed in a specific form or, where appropriate, in terms of a method or process for performing a disclosed function or for obtaining a disclosed result may be used to implement the various forms of the invention separately or in any combination of these features.

[0101] Although the present invention has been described in conjunction with the above exemplary embodiments, many equivalent modifications and variations will be apparent to those skilled in the art when the present disclosure is given. Therefore, the above exemplary embodiments of the present invention are considered to be illustrative rather than restrictive. Various changes may be made to the embodiments without departing from the spirit and scope of the present invention.

[0102] For the avoidance of any doubt, any theoretical explanations provided herein are intended to enhance the reader's understanding. The inventors do not wish to be bound by any of these theoretical explanations.

[0103] Any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.

[0104] Unless the context requires otherwise, throughout this specification, including the appended claims, the words "comprise" and "include" and variations such as "comprises", "comprising" and "including" will be understood to imply the inclusion of a stated integer or step or group of integers or steps without excluding any other integer or step or group of integers or steps.

[0105] It must be noted that the singular forms "a", "an", and "the" used in the specification and the appended claims include plural references unless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" one particular value and / or to "about" another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when a value is expressed as an approximation by using the antecedent "about", it is understood that the particular value forms another embodiment. The term "about" with respect to a value is optional and means, for example, + / - 10%.

[0106] When used herein, "and / or" should be taken as specific disclosure of each of the two specified features or components with or without the other. For example, "A and / or B" will be taken as specific disclosure of each of (i) A, (ii) B, and (iii) A and B, just as if each were individually listed herein.

Claims

1. A method for diagnosing a subtype of dementia in one or more patients, the method comprising: obtaining brain imaging data associated with the one or more patients; and The one or more patients are classified into a plurality of categories, including a first category of patients suffering from a first subtype of dementia and a second category of patients suffering from a second subtype of dementia, by providing the brain imaging data related to the one or more patients as input to a deep learning model, the deep learning model having been trained using brain imaging data from the plurality of patients, the brain imaging data including a first set of images showing evidence of temporal and parietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal areas instead of or in addition to the temporal and parietal areas, wherein the first set of images are labeled as being associated with the first subtype of dementia and the second set of images are labeled as being associated with the second subtype of dementia.

2. The method of claim 1, wherein the subtype of dementia is a subtype of Alzheimer's disease (AD), wherein the first subtype of dementia is AD, wherein the second subtype of dementia is AD combined with cerebrovascular disease or vascular dementia, wherein the second type of dementia is cerebrovascular disease or vascular dementia, wherein the second subtype of dementia is mixed AD and / or wherein the second subtype of dementia is not AD.

3. A method according to any preceding claim, wherein the classification is between a first class of patients suffering from a first subtype of dementia and a second class of patients suffering from a second subtype of dementia.

4. A method according to any preceding claim, wherein the deep learning model is a deep neural network classifier and / or wherein the deep learning model comprises a convolutional neural network (CNN), wherein the deep learning model comprises a model that has been pre-trained on unrelated image data, and / or wherein the deep learning model comprises a CNN that has been pre-trained using a deep residual learning framework.

5. A method according to any preceding claim, wherein the deep learning model comprises all layers of a CNN that have been pre-trained for image recognition except a classification layer, and a classification layer trained using the first and second sets of images and associated labels, wherein the classification layer optionally comprises a fully connected layer and a softmax layer.

6. A method according to any preceding claim, wherein the images in the first group and the second group show evidence of different metabolic activities in any one or more or all of the following: right frontal cortex, left frontal cortex, right temporal cortex, left temporal cortex, right parietal cortex, left parietal cortex, left cerebellum and right cerebellum.

7. A method according to any preceding claim, wherein the images in the second set of images show evidence of hypometabolism in one or more or all of the following regions of the brain, instead of or in addition to one or more of the following: right temporal cortex, left temporal cortex, right parietal cortex and left parietal cortex: right frontal cortex, left frontal cortex, left cerebellum and right cerebellum.

8. A method according to any preceding claim, wherein hypometabolism refers to a glucose uptake rate and / or blood flow and / or FDG-PET derived standardized uptake value ratio (SUVR) that is lower than expected from a control.

9. A method according to any preceding claim, wherein the brain imaging data is imaging data acquired using any functional brain imaging modality that provides information about the metabolic activity of the imaged brain region, optionally wherein the information about the metabolic activity of the imaged brain region is acquired by detecting glucose uptake in the imaged brain region and / or blood flow to the imaged brain region.

10. The method according to any preceding claim, wherein the brain imaging data is FDG-PET data or ASL data, and / or wherein the brain imaging data is baseline functional brain imaging data.

11. A method according to any preceding claim, wherein analyzing the data using a deep learning model comprises: analyzing a single section of brain imaging data for each patient and / or wherein the method comprises: selecting a single section of a set of brain imaging data for each patient, optionally wherein the single section is a single axial section at the level of the thalamus and / or wherein the single section is a section including at least a portion of the hippocampus and the entorhinal cortex.

12. A method according to any preceding claim, wherein the brain imaging data used to train the deep learning model comprises a single cross-section of a set of brain imaging data for each of a plurality of patients and / or wherein the method comprises: selecting a single cross-section of a set of brain imaging data for each of a plurality of patients, optionally wherein the single cross-section is a single axial cross-section at the level of the thalamus and / or wherein the single cross-section is a cross-section including at least a portion of the hippocampus and the entorhinal cortex.

13. A method according to any preceding claim, wherein the first set of images comprises one or more images for each of at least 30, at least 40, at least 50, at most 500, at most 200 or at most 100 patients, and / or the second set of images comprises one or more images for each of at least 30, at least 40, at least 50, at most 500, at most 200 or at most 100 patients, and / or the first set of images comprises one or more images for each of a number of patients who are within 10% or within 20% of the number of patients included in the one or more images in the second set of images, and / or wherein the first and second sets of images comprise one or more images for each of a plurality of patients, wherein the plurality of patients in the first and second sets of images are age-matched and gender-matched.

14. A method according to any preceding claim, wherein the first group of images has been labeled as being associated with a first subtype of dementia and the second group of images has been labeled as being associated with a second subtype of dementia through expert review of the first and second groups of images, wherein each image in the first group of images and the second group of images has been assigned the same label by at least two experts.

15. A method according to any preceding claim, wherein the brain imaging data used to train the deep learning model comprises images from a plurality of patients and images obtained from images from a plurality of patients by image enhancement, optionally wherein the image enhancement comprises creating a flipped version of one or more of the images and / or creating a randomly rotated version of one or more of the images.

16. A method according to any preceding claim, wherein the deep learning model classifies images into the first class and the second class with an accuracy of at least 90% or at least 95%, and / or wherein the deep learning model classifies images into the first class and the second class with a sensitivity of at least 90% or at least 94%, and / or wherein the deep learning model classifies images into the first class and the second class with a specificity of at least 90% or at least 95%.

17. A method according to any preceding claim, wherein the method further comprises: training the deep learning model using brain imaging data from a plurality of patients, the brain imaging data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than the temporal and parietal regions instead of or in addition to the temporal and parietal regions, wherein the first set of images are labeled as being associated with a first subtype of dementia and the second set of images are labeled as being associated with a second subtype of dementia.

18. A method of providing a tool for diagnosing a subtype of dementia in one or more patients, the method comprising: Acquiring training imaging data comprising a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in areas of the brain other than or in addition to the temporal and parietal areas, wherein the first set of images are labeled as being associated with a first subtype of dementia and the second set of images are labeled as being associated with a second subtype of dementia; and A deep learning model is trained using the training imaging data to classify patients into a plurality of categories, including a first category of patients having a first subtype of dementia and a second category of patients having a second subtype of dementia.

19. A method for selecting an individual who has been diagnosed with AD or is likely to have AD to participate in a clinical trial, the method comprising: Classifying the individual into a plurality of categories, including a first class of patients having a first subtype of dementia and a second class of patients having a second subtype of dementia, by providing brain imaging data associated with the individual as input to a deep learning model, the deep learning model having been trained using brain imaging data from a plurality of patients, the brain imaging data including a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in regions of the brain other than or in addition to the temporal and parietal regions, wherein the first set of images are labeled as being associated with the first subtype of dementia and the second set of images are labeled as being associated with the second subtype of dementia; and The individual is selected or excluded from participation in the clinical trial depending on whether the individual is classified as having a first dementia subtype or a second dementia subtype.

20. A method for providing a prognosis for an individual who has been diagnosed with AD or is likely to have AD, the method comprising: Classifying the individual into a plurality of categories, including a first class of patients having a first subtype of dementia and a second class of patients having a second subtype of dementia, by providing brain imaging data associated with the individual as input to a deep learning model, the deep learning model having been trained using brain imaging data from a plurality of patients, the brain imaging data including a first set of images showing evidence of temporoparietal hypometabolism and a second set of images showing evidence of hypometabolism in regions of the brain other than or in addition to the temporal and parietal regions, wherein the first set of images are labeled as being associated with the first subtype of dementia and the second set of images are labeled as being associated with the second subtype of dementia; and Determining a prognosis for the individual based on the individual being classified as having the first subtype of dementia or the second subtype of dementia, as appropriate, wherein determining the prognosis includes determining that if the individual is classified as the second type, the individual is likely to have a faster rate of cognitive decline than if the individual is classified as the first type.

21. A system for diagnosing a subtype of dementia and / or providing a tool for diagnosing a subtype of dementia, the system comprising: one or more processors and a computer-readable memory storing instructions, the instructions causing the processor to execute a method according to any one of claims 1 to 18, wherein the system further comprises a data acquisition device configured to acquire brain imaging data associated with one or more patients.

22. A non-transitory computer readable storage medium containing machine executable instructions which, when executed on a processor, cause the processor to perform the method according to any one of claims 1 to 18.

23. A computer program comprising executable codes which, when run on a computer, cause the computer to perform the method according to any one of claims 1 to 18.

Images