Methods, systems and uses for determining abdominal aortic calcification

Abstract

Methods, systems, and uses for determining abdominal aortic calcification involve determining abdominal aortic calcification from lateral lumbar images and diagnostic / prognostic uses for abdominal aortic calcification. The method for determining abdominal aortic calcification includes receiving lateral lumbar images and determining, using one or more processors, a score representative of abdominal aortic calcification from the images, the determining step including encoding the images to identify visual features and decoding the visual features to calculate a plurality of calcification scores, each for a segment of the abdominal aorta.

Description

[Technical Field]

[0001]

[0001] The present invention relates to determining abdominal aortic calcification from lateral lumbar images and the diagnostic / prognostic uses of abdominal aortic calcification. [Background technology]

[0002] The following discussion of the background art is intended solely to facilitate an understanding of the present invention. It should be understood that this discussion is not an acknowledgment or admission that any of the material referred to was part of the common general knowledge as of the priority date of this application.

[0003]

[0003] Cardiovascular disease (CVD) is a leading cause of death worldwide and a significant contributor to disability worldwide. Vascular calcification is a marker of subclinical CVD and occurs when calcium accumulates within the walls of arteries undergoing the atherosclerotic process. Calcification often begins decades before a clinical event such as a heart attack or stroke occurs. The abdominal aorta is one of the first vascular beds to exhibit calcification. The presence and extent of abdominal aortic calcification (AAC) is associated with an increased risk of future cardiovascular hospitalization and mortality.

[0004] The extent and severity of AAC can be assessed using lateral lumbar radiographs, lateral lumbar fracture assessment (VFA), dual-energy x-ray absorptiometry (DXA), and quantitative computed tomography (QCT). VFA and DXA scans have minimal radiation but lower resolution. These scans can be used to semi-quantify AAC using the widely adopted Kauppila 24-point scoring method (AAC24), which measures calcification along the length of the abdominal aorta from L1 to L4. However, obtaining manual assessments of DXA images is not only time-consuming and expensive, but also subjective.

[0005]

[0005] The AAC24 scoring system scores the AAC for each vertebral level (L1-L4) and scores both the anterior and posterior aortic walls as 0 (no calcification), 1 (≦1 / 3 of the aortic wall), 2 (>1 / 3 to ≦2 / 3 of the aortic wall), or 3 (>2 / 3 of the aortic wall), giving a maximum possible score of 24.

[0006]

[0006] AAC severity is typically classified as low (AAC24 score of 0 or 1), moderate (AAC24 score of 2-5), and high (AAC24 score of 6 or greater).

[0007] Some preliminary work has been done to automatically predict the overall AAC24 score of radiological scans using machine learning models (Chaplin, L., Cootes, T.: Automated scoring of aortic calcification in vertebral fracture assessment images. In: Medical Imaging 2019: Computer-Aided Diagnosis, vol. 10950, pp. 811-819. SPIE (2019); Elmasri, K., Hicks, Y., Yang, X., Sun, X., Pettit, R., Evans, W.: Automatic detection and quantification of abdominal aortic calcification in dual energy x-ray absorptiometry. Procedia Computer Science 96, 1011-1021 (2016); and Reid, S., Schousboe, JT, Kimelman, D., Monchka, BA, Jozani, MJ, Leslie, WD: Machine learning for automated abdominal aortic calcification scoring of DXA vertebral fracture assessment images: A pilot study. Bone 148, 115943 (2021)). However, these techniques only produce an overall AAC24 score, which has limited reproducible accuracy.

[0008]

[0008] Furthermore, the human-generated AAC24 score is currently used only as an indicator or prognosis of CVD.

[0009]

[0009] The present invention has been made in view of the above background.

[0010]

[0010] Throughout this specification, unless the context requires otherwise, the word "comprise" or variations such as "comprises" or "comprising" will be understood to imply the inclusion of a stated integer or group of integers, but not the exclusion of any other integer or group of integers.

[0011]

[0011] Throughout this specification, unless the context requires otherwise, the word "include" or variations such as "includes" or "including" will be understood to imply the inclusion of a stated integer or group of integers, but not the exclusion of any other integer or group of integers. Summary of the Invention

[0012] According to one aspect of the present invention, there is provided a method for determining abdominal aortic calcification. And , including the abdominal aortic section of the patient's abdominal aorta Receiving a lateral lumbar image wherein the abdominal aorta section comprises a plurality of abdominal aorta segments. and, Using one or more processors lateral lumbar region and determining a score representative of abdominal aortic calcification from the image, wherein the determining step comprises: One or more processors may select a lateral lumbar image in the abdominal aorta section. To identify visual features, lateral lumbar region encoding an image; one or more processors implementing a forward decoder that decodes the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; the one or more processors implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score associated with a posterior portion of one abdominal aorta segment; A method is provided in which each of the forward and backward decoders is a trained machine learning decoder, and the forward decoder is trained separately from the backward decoder.

[0013] In one embodiment, Each abdominal aorta The segments are Patient L1, L2, L3 and L4 vertebrae One of Corresponds to.

[0014] In one embodiment, The encoding step includes identifying the L1, L2, L3, and L4 abdominal aortic segments in the lateral lumbar image. Includes.

[0015] In one embodiment, Each of the forward and backward decoders includes a CNN, and the method includes training the forward and backward decoders using a long short-term memory to store the sequence of abdominal aortic segment calcification scores and an attention module. Includes.

[0016] In one embodiment, The encoding step includes providing the lateral lumbar images to a trained convolutional neural network (CNN); extracting visual features using the convolutional neural network (CNN); Includes.

[0017] In one embodiment, The visual features are obtained from the last convolutional layer of a convolutional neural network (CNN).

[0018] In one embodiment, Calcification scores are classified into risk categories.

[0019] In one embodiment, The risk category is used to distinguish the prognosis of the disease.

[0020] In one embodiment, The risk category includes risk of CVD.

[0021] In one embodiment, The method includes comparing the risk of CVD to a threshold and, when the threshold is exceeded, classifying the patient associated with the lateral lumbar image as being at risk of CVD.

[0022] In one embodiment, The risk category includes risk of falls and / or fractures later in life.

[0023] In one embodiment, The method includes comparing the risk of later-life falls and / or fractures to a threshold, and if the threshold is exceeded, classifying the patient associated with the lateral lumbar image as having a risk of later-life falls and / or fractures.

[0024] In one embodiment, Risk categories include risk of late-life osteoporosis or osteoporotic fractures.

[0025] In one embodiment, The method includes comparing the risk of osteoporosis to a threshold, and classifying the patient associated with the lateral lumbar image as being at risk for osteoporosis when the threshold is exceeded.

[0026] In one embodiment, Risk categories include risk of late-life dementia.

[0027] In one embodiment, the method further comprises the step of comparing the risk of dementia to a threshold. and , threshold When the patient of Classify as at risk of dementia The method includes the steps of:

[0028] In one embodiment, Risk Category This includes the risk of diabetes.

[0029] In one embodiment, The method includes comparing the risk of diabetes to a threshold and, when the threshold is exceeded, classifying the patient associated with the lateral lumbar image as being at risk of diabetes.

[0030]

[0030] According to one aspect of the present invention, there is provided a system for determining abdominal aortic calcification, comprising: at least one processor for determining a score representative of abdominal aortic calcification from received lateral lumbar images of the patient's abdominal aorta, the lateral lumbar images including an abdominal aortic section having a plurality of abdominal aortic segments; At least one processor Implementing an encoder that identifies visual features in the lateral lumbar image in the abdominal aorta section; implementing an anterior decoder that decodes the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score being associated with one abdominal aorta segment; A system is provided in which each of the forward and backward decoders is a trained machine learning decoder, and the forward decoder is trained separately from the backward decoder.

[0031] In one embodiment, Each abdominal aortic segment corresponds to one of the patient's L1, L2, L3 and L4 vertebrae.

[0032] In one embodiment, The encoder is configured to identify the L1, L2, L3, and L4 abdominal aorta segments in the lateral lumbar images.

[0033] In one embodiment, The system includes an imager for generating a lateral lumbar image.

[0034] In one embodiment, The at least one processor is configured to perform image cropping, and the system includes a resizer configured to crop and resize the lateral lumbar images to a predetermined size and resolution for encoding according to the source of the lateral lumbar images.

[0035] In one embodiment, The encoder includes a trained convolutional neural network (CNN) to extract visual features from the lateral lumbar images. The final convolutional layer can output the visual features.

[0036] In one embodiment, Each of the forward and backward decoders includes a trained CNN, which may include a long short-term memory network and an attention module.

[0037] In one embodiment, Each of the forward and backward decoders includes a global pooling layer, a dense layer with normalized linear activation units (Relu activations), and a dense layer with linear activations.

[0038] [0037a] In one embodiment, the at least one processor is configured to implement a risk category classifier for determining a risk category of a disease.

[0039] According to one aspect of the present invention, there is provided a method for diagnosing a disease, the method comprising: receiving a lateral lumbar image including an abdominal aortic section of the patient's abdominal aorta, the abdominal aortic section including a plurality of abdominal aortic segments; and determining, using one or more processors, a score representative of abdominal aortic calcification from the lateral lumbar images, wherein the determining step comprises: one or more processors encoding the lateral lumbar images to identify visual features within the images in an abdominal aorta segment; one or more processors implementing an anterior decoder that decodes the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; the one or more processors implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score associated with a posterior portion of one abdominal aorta segment; each of the forward and backward decoders is a trained machine learning decoder, the forward decoder being trained separately from the backward decoder; The method includes determining a disease risk category based on the determined plurality of anterior and posterior calcification scores.

[0040] In one embodiment, the method includes a step of identifying a prognosis of the disease from the determined risk category.

[0041]

[0040] Also Also disclosed is the use of abdominal aortic calcification determined by a method comprising: receiving a lateral lumbar image; and determining, using one or more processors, a score representative of abdominal aortic calcification from the image, wherein the determining step comprises: encoding the image to identify visual features; decoding the visual features to calculate a plurality of calcification scores, each for a segment of the abdominal aorta; The use includes determining a risk category for a disease based on the determined abdominal aortic calcification.

[0042] In one embodiment, the use comprises the step of identifying a disease prognosis from the determined risk category.

[0043]

[0042] A method for diagnosing or prognosing a disease or condition in a human is also disclosed, which method includes a step of classifying into risk categories a calcification score obtained by the method for determining abdominal aortic calcification according to the present invention as described herein.

[0044]

[0043] In one embodiment of the method for diagnosing or prognosing a disease or health risk in a human, the disease comprises CVD, diabetes, dementia, or osteoporosis.

[0045] In one embodiment of the method for diagnosing or prognosing a disease or health risk in a human, the health risk comprises a risk of falls and / or fractures later in life.

[0046]

[0045] A screening test for diagnosing or prognosing a disease or condition in a human is also disclosed, which includes a step of classifying a calcification score obtained by the method for determining abdominal aortic calcification according to the present invention as described herein into a risk category.

[0047] In one embodiment of a screening test for diagnosing or prognosing a disease or health risk in a human, the disease comprises CVD, diabetes, dementia, or osteoporosis.

[0048] In one embodiment of a screening test for diagnosing or prognosing a disease or health risk in a human, the health risk comprises risk of falls and / or fractures later in life.

[0049] Also disclosed is the use of the method for determining abdominal aortic calcification according to the invention described herein in the diagnosis or prognosis of disease or health risk in humans.

[0050]

[0049] In one embodiment of the use of the method for determining abdominal aortic calcification according to the present invention described herein, the disease comprises CVD, diabetes, dementia, or osteoporosis.

[0051]

[0050] In one embodiment of the use of the method for determining abdominal aortic calcification according to the present invention described herein, the health risk comprises the risk of falls and / or fractures later in life.

[0052]

[0051] Also disclosed is a method for treating a human having a disease including CVD, diabetes, dementia, or osteoporosis after diagnosing or prognosing the disease, the method comprising a step of classifying into risk categories the calcification score obtained by the method for determining abdominal aortic calcification according to the present invention described herein.

[0053]

[0052] Also A method for treating a person with a health risk, including a risk of falls and / or fractures in later life, after diagnosing or predicting a prognosis for the health risk, comprising classifying into risk categories a calcification score obtained by the method for determining abdominal aortic calcification according to the present invention described herein. Also disclosed will be done.

[0054]

[0053] Also a program for controlling one or more processors, the program comprising instructions stored on a non-volatile medium that controls the processors to perform any one of the methods described herein or to operate as any one of the systems described herein. Also disclosed will be done.

[0055] In order to provide a better understanding, embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0056] [Figure 1]FIG. 1 is a schematic diagram of a system for determining abdominal aortic calcification score, according to one embodiment of the present invention. [Figure 2] FIG. 2 is a schematic block diagram of a processor of the system of FIG. 1. [Figure 3] FIG. 3 is a schematic block diagram of a decoder of the processor of FIG. 2; [Figure 4] FIG. 4 is a schematic functional diagram of a sub-decoder of the decoder of FIG. 3; [Figure 5] FIG. 5 is a schematic functional diagram of the attention module of each of the sub-decoders of FIG. 4. [Figure 6] FIG. 1 is a schematic diagram of a method for determining an abdominal aortic calcification score, according to one embodiment of the present invention. [Figure 7] 1 is a scatter plot of the predictive accuracy of an exemplary system according to an embodiment of the present invention and another system for determining abdominal aortic calcification score; DETAILED DESCRIPTION OF THE INVENTION

[0057]

[0055] Referring to FIG. 1, a system 10 for determining an abdominal aortic calcification score 16 from a lateral lumbar image 12 is shown, including a neural network processor 14. The image 12 may be a lateral lumbar radiograph obtained from an imager for generating the lateral lumbar image. For example, the imager may be a lateral lumbar fracture assessment (VFA), dual-energy X-ray absorptiometry (DXA), or quantitative computed tomography (QCT) device. Other image sources, such as standard X-ray images and bone density images taken from a perspective that includes the relevant lumbar segment, may also be used. In this embodiment, the image 12 should show the length of the abdominal aorta adjacent to each lumbar segment of the L1, L2, L3, and L4 vertebrae in a human. Other aortic segments (including those adjacent to the lower lumbar region) may be used in other mammals that experience calcification. The score 16 can be used to diagnose or prognose disease or health risk in humans. The disease may include one or more of CVD, diabetes, dementia, or osteoporosis. CVD includes various types of heart disease, stroke, and vascular disease.

[0058]

[0056] Referring to Figure 2, the neural network processor 14 comprises a first sub-processor module operating as an encoder 22 for identifying visual features within the image 12 and a second sub-processor module operating as a decoder 24 for calculating calcification scores, each score for a segment of the abdominal aorta adjacent to each lumbar segment.

[0059] Preferably, the image 12 is pre-processed by a pre-processor 26 before the encoder 22 identifies visual features. Preferably, the calculated calcification scores from the decoder 24 are provided to a user by an output 28. Preferably, the output 28 combines the calculated calcification scores into an overall score for output to the user. Preferably, the output 28 applies analysis to the calculated calcification scores and / or the overall score and provides the analysis to the user.

[0060] In one embodiment, pre-processor 26 is configured to crop and resize image 12 according to the image's source to a predetermined size and resolution for encoding. In one embodiment, pre-processor 26 comprises an image cropper and an image resizer. In one embodiment, the cropping performed by the image cropper includes applying one or more affine transformations. In one embodiment, the resizing performed by the image resizer uses nearest neighbor interpolation.

[0061] In one embodiment, the encoder 22 includes a trained convolutional neural network (CNN) for extracting visual feature maps from the (preferably cropped and resized) images. In one embodiment, the CNN is a residual network, including a deep neural network with a residual block. In the residual block, direct connections skip some layers of the neural network. In one embodiment, the final convolutional layer outputs the visual feature map, rather than subsequent classification layers of the CNN. In one embodiment, the CNN is trained using stochastic gradient descent using corresponding identified segments of the abdominal aorta in the lateral lumbar images and the lateral lumbar images. The classification layer in the trained CNN outputs identification of segments in each image, but in this embodiment, it is used only for training purposes.

[0062] 3, in one embodiment, decoder 24 comprises at least two sub-decoders, preferably a forward decoder 32 for decoding the anterior wall of the abdominal aorta as an anterior calcification score adjacent each of the lumbar segments, and a backward decoder 34 for decoding the posterior wall of the abdominal aorta as an anterior calcification score adjacent each of the lumbar segments. In one embodiment, the sub-decoders (forward decoder 32 and backward decoder 34) are independently trained classifiers.

[0063] 4, in one embodiment, each sub-decoder 32, 34 includes a trained neural network. In one embodiment, each sub-decoder 32, 34 is trained using a long short-term memory network for storing a sequence of segmented scores and an attention module for generating the segment scores for the respective sub-decoder 32, 34.

[0064]

[0062] In one embodiment, each decoder includes a global pooling layer followed by a dense layer with a normalized linear activation unit (Relu activation) and another dense layer with linear activation.

[0065] Each sub-decoder (a) and (b) in FIG. 4 independently maximizes the log-likelihood over the parameter space.

number

[0066]

[0064] The log-likelihood of the joint probability distribution log p(y|V;θ) can be decomposed as follows:

number

[0067] The long short-term memory (LSTM) module generates y, and therefore the conditional probability log p(y|V) (omitting θ for convenience) is log(py t |y 1 ,…y t-1 ,V)=g(h t ,g t ), where g is a nonlinear function and c t is the context vector, and h tis the hidden state of the LSTM at time t. t is, h t =LSTM(s t ,h t-1 ,m t-1 ), where s t is the input vector, and h t-1 and m t-1 is the hidden state and memory cell vector at time t-1.

[0068] Referring to FIG. 5, the context vector c t An attention module is used to compute c, so that the context depends on the specific region in the image (via the image feature map) as well as the sub-decoder output. t is c t =q(V,h t-1 ), where q is the attention function and h t is the hidden state of the LSTM at time t. The distribution of attention across feature maps V (corresponding to different regions of the image) is calculated using a feedforward network, z t =W a tanh(W v V+W h h t-1 ) and β t = softmax(z t ), where W a , W v , and W h are the learnable parameters, and β are the attention weights over the feature map V. Finally, c t can be calculated as follows:

number

[0069]

[0067] The model is trained using weighted cross-entropy loss, where the weight of each class is set based on the data distribution.

[0070] The two sub-decoders are trained independently to maximize the objective function given in Equation 2.

[0071] In one embodiment, output 28 is configured to combine each of the forward L1, L2, L3, and L4 scores into a combined forward score and each of the backward L1, L2, L3, and L4 scores into a combined backward score. In one embodiment, the output is configured to combine the combined forward score and the combined backward score into an overall score.

[0072] In an alternative embodiment, output 28 is configured to combine each of the anterior and posterior calcification scores for each of L1, L2, L3, and L4 into a combined L1, L2, L3, and L4 score. In one embodiment, the method includes combining each of the combined L1, L2, L3, and L4 scores into an overall score.

[0073]

[0071] Referring now to Figure 6, the method of use and operation of the system 10 for determining abdominal aortic calcification will now be described.

[0074] At 42, a lateral lumbar image 12 is received from the imager. A processor 14 determines a score 16 representing abdominal aortic calcification from the image 12. A pre-processor 26 crops and resizes the image to a predetermined size and resolution for encoding according to the source of the image.

[0075] At 44, the encoder 22 extracts the visual feature map. At 46 and 50, each sub-decoder 32, 34 decodes the visual feature map into a set of calcification scores, which are a set of anterior calcification scores 48 and a set of posterior calcification scores 52. The output 28 provides a combined L1-L4 set of scores 12 (54). The output 28 may also provide a combined AAC24 score from the combined L1-L4 set of scores 12.

[0076] In one embodiment, output 28 is configured to classify the combined L1-L4 set of scores 12 (or the combined AAC24 score) into a risk category.

[0077] In one embodiment, output 28 is configured to classify risk categories to identify disease prognosis.

[0078] In one embodiment, the risk category includes a risk of CVD. In one embodiment, the method includes comparing the risk of CVD to a threshold, and if the threshold is exceeded, the patient of the image is classified as having a risk of CVD. The classification of risk of CVD can be used as a screening tool for referral to a cardiologist.

[0079] In one embodiment, the risk score includes risk of later-life falls and / or fractures. In one embodiment, the method includes comparing the risk of later-life falls and / or fractures to a threshold, and if the threshold is exceeded, the subject of the patient image is classified as having risk of later-life falls and / or fractures. Classification of later-life falls and / or fractures risk can be used as a screening tool for referral for intervention.

[0080] In one embodiment, the risk score includes a risk of later-life osteoporosis. In one embodiment, the method includes comparing the risk of osteoporosis to a threshold, and if the threshold is exceeded, the patient of interest in the image is classified as having a risk of osteoporosis. Classification of later-life osteoporosis risk can be used as a screening tool for referral for intervention.

[0081] In one embodiment, the risk score includes a risk of late-life dementia. In one embodiment, the method includes comparing the risk of dementia to a threshold, and if the threshold is exceeded, the patient subject of the image is classified as having a risk of dementia. The classification of risk of late-life dementia can be used as a screening tool for referral for intervention.

[0082] In one embodiment, the risk score includes a risk of later-life diabetes. In one embodiment, the method includes comparing the risk of diabetes to a threshold, and if the threshold is exceeded, the patient of the image is classified as having a risk of diabetes. The classification of later-life diabetes risk can be used as a screening tool to recommend interventions and / or medications and / or dietary changes and / or lifestyle changes.

[0083] Example The dataset consisted of 1,916 randomly selected lateral lumbar scans from a bone density machine, acquired using an iDXA GE machine with a resolution of at least 1600 x 300 pixels, and was used to train the system. The disease severity distribution of the 1,916 scans was 829 low-risk, 445 moderate-risk, and 642 high-risk. These scans were annotated with expert-annotated AAC24 scores, but the location of calcification pixels was not annotated on the scans. The dataset with a distribution of zero scores is highly asymmetric relative to L1 and L2, likely because vascular calcification typically begins near L4 and L3 and then progresses upward. The anterior segments in the dataset had 176 unique (out of 44 = 256 possible) combinations, but only 29 of the combinations occurred more than 10 times. The most frequent sequence was [0,0,0,0], which occurred 904 times, followed by [0,0,0,1], which occurred 77 times. For the posterior segment, the dataset had 190 unique combinations, of which only 30 appeared more than 10 times. Again, [0,0,0,0] was the most frequent combination, appearing 786 (41%) times.

[0084]

[0082] The distribution of scores in the dataset was as follows: [Table 1]

[0085] The preprocessor cropped each scan by 50% from the top, 40% from the left, and 10% from the right. The cropped images were resized to 900 x 300 pixels using nearest neighbor interpolation and rescaled to values ​​between 0 and 1.

[0086] The training dataset was augmented by applying various affine transformations to the images, such as translation [+20,-20], scaling [+20,-20], shear [0.01°,0.05°], and rotation [+10°,-10]. The TorchVision library for data augmentation and the PyTorch machine learning library were used for model training and evaluation.

[0087] Resnet152v2 was pre-trained on ImageNet to be used as the encoder 22, but other models may be used. Feature maps from the last convolutional layer of the pre-trained CNN, without using the classification layer. For an input image size of 900x300, the extracted feature maps are 29x10x2048 in size. The feature maps are flattened to 290x2048 and then passed to the decoder. ant and decoder post The two signals are fed separately to two decoding networks called

[0088]

[0086] Two decoders, decoder ant and decoder postwere trained independently using the sequences of ground truth scores for the anterior and posterior segments, respectively. Furthermore, after training was completed, the output scores of both decoders (for a given test image) were summed to obtain a single score corresponding to each lumbar region. Finally, the scores for L1 through L4 were summed to obtain the AAC24 score.

[0089] Both decoders consisted of LSTMs with a hidden size of 512, based on an attention module, and an output sequence length of 4. A 10-fold stratified cross-validation was performed (the data was split based on the distribution of AAC24 scores, so that this distribution was maintained across all splits). In each fold, 1,724 examples were used to train the network, and 192 were used for validation. Early stopping was based on the average Pearson correlation between predicted and ground truth segment scores. Dropout was used as a regularization strategy (first after the hidden layer of the LSTM (alpha = 0.5), then another layer (alpha = 0.4) was used before the final FC layer).

[0090] The resulting scores were evaluated by summing all individual granularity scores using the same data set. Compared with human assessment, classification of patients into three risk categories, low, medium, and high, had an accuracy, sensitivity, and specificity of 82%, 74%, and 80%, respectively, for the test set. The AAC24 scores generated by the present invention were highly correlated (>80%) with human assessment.

[0091] The pipeline model of Reid et al. (mentioned in the background above) fgs (M base Implemented (with minor modifications) as M baseIn

[2016] , a baseline CNN was trained using Resnet152v2 as its encoder. The decoder consists of a global pooling layer followed by a dense layer with Relu activation and another dense layer with linear activation. The generated AAC24 scores are classified into three risk levels based on a risk threshold.

[0092] Baseline M in the one-vs-rest setting using cumulative AAC24 prediction scores base (NPV is the negative predictive value, and PPV is the positive predictive value) fgs The performance comparison of the models is as follows: [Table 2]

[0093]

[0091] M fgs The mean classification accuracy of 81.98 + / - 2.5% is significantly better than the baseline accuracy of 70.77 + / - 3.2%. fgs The average three-class classification accuracy is 72.8±2.9%, while that of the baseline is 55.8±3.2%. Therefore, M fgs The model predicts AAC24 scores more accurately compared to the baseline model.

[0094]

[0092] M fgs We compare the single AAC24 scores (for all lumbar regions) output from M with the corresponding ground truth scores in the scatter plot of Figure 7 and the confusion matrix in the table below. fgs The model is very good at classifying low-risk and high-risk patients. This figure provides evidence that fine-grained scoring results in significantly (p<<0.01) better prediction and higher correlation with human scores. [Table 3] [Table 4]

[0095]

[0093] We also confirmed whether predicting AAC scores in two segments is better than predicting AAC scores horizontally across each lumbar region, e.g., L1 or L2, by training a variant of the model with a single decoder to predict a sequence of scores for each lumbar region, L1 through L4, where the score for L1 is the sum of the scores for L1 and L2. ant and L1 post is the sum of

[0096]

[0094] Variant model (M fgs ) was used to determine whether predicting AAC scores in two segments was better than predicting AAC scores horizontally across each lumbar region, e.g., L1 or L2. M fgs A single decoder is used to predict the sequence of scores for each lumbar region L1-L4. fgs trained as variants of the model, L1 score is L1 ant and L1 post The AAC score is the sum of the predicted horizontal AAC scores across each vertebra and the predicted AAC scores across each segment (anterior and posterior). Between predicting the score vertically and Comparison of, i.e., M fgs and M fgs A comparison between is shown in the table below. [Table 5]

[0097]

[0095] M fgs The correlation between the scores predicted by and the human annotated scores of our variants M fgs This is significantly better than the correlation provided by (p<0.01).

[0098] Using the Score

[0096] The output 28 of the system 10 may be configured to identify, from the risk category, a risk of developing a disease or condition or a predicted prognosis, which may be used as a screening tool for further investigation of the predicted prognosis by a relevant specialist and / or to provide remedial / preventive treatment or in diagnosis.

[0099]

[0097] The CVD risk score is as described above. Preventive treatments for the predicted prognosis of CVD include increasing fruit and vegetable consumption, improving diet, reducing sedentary time, and / or increasing physical activity.

[0100]

[0098] An osteoporosis risk score can be determined from the AAC24 score, for example, the combined AAC24 score can be calculated based on hip bone mineral density (rs = 0.077 , p = 0.013) and inversely correlated with heel broadband ultrasound attenuation (rs = -0.074, p = 0.020) and stiffness index (rs = -0.073, p = 0.022). Severe AAC was associated with a higher likelihood of common fractures and hip injuries. Moderate-to-severe AAC (AAC24 score > 1) had an increased fracture risk compared with low AAC (HR 1.48 [1.15-1.91], p = 0.002; HR 1.46 [1.07-1.99], p = 0.019, respectively).

[0101]

[0099] Prognostic preventative treatments for osteoporosis include oral calcium supplements.

[0102] A risk score for fall-related hospitalization can be determined from the AAC24 score attributable to weak grip strength. In one embodiment, an AAC24 score indicating risk for fall-related hospitalization is at least 2, 3, 4, 5, or 6. In one embodiment, an AAC24 score of at least 6, 7, 8, 9, or 10 presents a strong indicator of risk for fall-related hospitalization. In the study, over 14.5 years, 413 women (39.2%) experienced a fall-related hospitalization. Using a multivariate-adjusted model, each unit increase in baseline AAC24 was associated with a 3% increase in the relative hazard of fall-related hospitalization (HR 1.03 95% CI, 1.01-1.07). Compared with women without AAC, women with any AAC had a 40% (HR 1.40 95% CI, 1.11-1.76) and 39% (HR 1.39 95% CI, 1.10-1.76) higher risk of fall-related hospitalization in the minimal and multivariate-adjusted models, respectively.

[0103]

[0101] Furthermore, the presence of AAC is associated with more than 7 in 10 women experiencing hospitalization related to falls, a 39% higher risk compared to women without AAC.

[0104]

[0102] Preventive treatment of the predicted outcomes of fall-related hospitalization includes fall prevention programs that include strength training.

[0105] A dementia risk score can be determined from the AAC24 score. In one embodiment, an AAC24 score indicating dementia risk is at least 2, 3, 4, 5, or 6. In one embodiment, an AAC24 score of at least 6, 7, 8, 9, or 10 is a strong indicator of dementia risk. At baseline, women were 75.0 + / - 2.6 years old; 44.7% had low AAC, 36.4% had moderate AAC, and 18.9% had severe AAC. Over 14.5 years, 150 women (15.7%) had late-life dementia hospitalization (n = 132) and / or death (n = 58). Compared with women with low AAC, women with moderate and severe AAC were more likely to experience late-life dementia hospitalization (9.3%, 15.5%, and 18.3%, respectively) and death (2.8%, 8.3%, and 9.4%, respectively). After adjusting for cardiovascular risk factors and APOE, women with moderate and severe AAC had a twofold relative hazard of late-life dementia compared with women with low AAC (moderate, aHR 2.03 95% CI 1.38-2.97; severe, aHR 2.10 95% CI 1.33-3.32).

[0106]

[0104] Prognostic and preventative treatments for dementia include lifestyle changes and drug therapy.

[0107] A diabetes risk score can be determined from the AAC24 score. In one embodiment, the diabetes is type I, alternatively type II, or alternatively both type I and type II. In one embodiment, the AAC24 score, which indicates future diabetes risk, is at least 2, 3, 4, 5, or 6. In one embodiment, an AAC24 score of at least 6, 7, 8, 9, or 10 is a strong indicator of future diabetes risk. In the study, AAC was more prevalent in patients with diabetes mellitus (DM), with an AAC prevalence of 29% in DM men (n=70) vs. 17% in non-DM men (n=62) (p=0.05), and 26% vs. 19% in DM women (n=63) vs. non-DM women (n=82) (p=0.06).

[0108]

[0106] Preventive treatment of the expected outcomes of diabetes includes lifestyle changes and medication.

[0109]

[0107] The present invention not only overcomes the bottleneck of manual AAC24 determination, but also provides improved results in continuous "fine-grained" scoring and a more accurate derived overall score, which can be used for the diagnosis and / or prognosis of several diseases.

[0110]

[0108] Modifications can be made to the invention within the scope of what has been described and shown in the drawings, and such modifications are intended to form a part of the invention as described herein.

Claims

1. 1. A method for determining abdominal aortic calcification, comprising: receiving a lateral lumbar image including an abdominal aortic section of the patient's abdominal aorta, the abdominal aortic section including a plurality of abdominal aortic segments; and determining, using one or more processors, a score representative of abdominal aortic calcification from the lateral lumbar images, wherein said determining step comprises: the one or more processors encoding the lateral lumbar images to identify visual features within the lateral lumbar images in the abdominal aorta section; the one or more processors implementing a forward decoder that decodes the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; the one or more processors implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score associated with a posterior portion of one abdominal aorta segment; 10. A method wherein each of the forward and backward decoders is a trained machine learning decoder, the forward decoder being trained separately from the backward decoder.

2. each abdominal aortic segment corresponds to one of the patient's L1, L2, L3, and L4 vertebrae; The method of claim 1 , wherein the encoding step includes identifying L1, L2, L3, and L4 abdominal aortic segments in the lateral lumbar image.

3. 2. The method of claim 1 , wherein each of the forward and backward decoders comprises a CNN, the method comprising training the forward and backward decoders using a long short-term memory to store a sequence of abdominal aorta segment calcification scores and an attention module.

4. 4. The method of claim 1, wherein the encoding step comprises providing the lateral lumbar images to a trained convolutional neural network (CNN) and extracting visual features using the convolutional neural network (CNN).

5. The method of claim 4 , wherein the visual features are obtained from a last convolutional layer of the convolutional neural network (CNN).

6. The method of claim 1 , wherein the calcification score is classified into risk categories.

7. The method of claim 6 , wherein the risk category is used to distinguish disease prognosis.

8. 7. The method of claim 6, wherein the risk categories include risk of CVD, risk of falls and / or fractures in later life, risk of osteoporosis or osteoporotic fractures in later life, risk of dementia in later life, and / or risk of diabetes.

9. 1. A system for determining abdominal aortic calcification, comprising: at least one processor for determining a score representative of abdominal aortic calcification from received lateral lumbar images of the patient's abdominal aorta, the lateral lumbar images including an abdominal aortic section having a plurality of abdominal aortic segments; the at least one processor: implementing an encoder that identifies visual features within the lateral lumbar image in the abdominal aorta section; implementing an anterior decoder that decodes the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score being associated with one abdominal aorta segment; A system wherein each of the forward and backward decoders is a trained machine learning decoder, and the forward decoder is trained separately from the backward decoder.

10. each abdominal aortic segment corresponds to one of the patient's L1, L2, L3, and L4 vertebrae; The system of claim 9 , wherein the encoder is configured to identify L1, L2, L3, and L4 abdominal aorta segments in the lateral lumbar images.

11. 11. The system of claim 9 or 10, wherein the encoder comprises a trained convolutional neural network (CNN) for extracting visual features from the lateral lumbar images.

12. The system of claim 11 , wherein a final convolutional layer outputs the visual features.

13. 10. The system of claim 9, wherein the forward and backward decoders each comprise a trained CNN.

14. 14. The system of claim 13, wherein the forward and backward decoders each comprise a long short-term memory network and an attention module.

15. 14. The system of claim 13, wherein each of the forward and backward decoders includes a global pooling layer, a dense layer with normalized linear activation units (Relu activations), and a dense layer with linear activations.

16. 10. The system of claim 9, wherein the at least one processor is configured to implement a risk category classifier for determining a risk category of a disease.

17. 1. A method for diagnosing a disease, said method comprising: receiving a lateral lumbar image including an abdominal aortic section of the patient's abdominal aorta, the abdominal aortic section including a plurality of abdominal aortic segments; and determining, using one or more processors, a score representative of abdominal aortic calcification from the lateral lumbar images, wherein said determining step comprises: the one or more processors encoding the lateral lumbar images to identify visual features within the images in the abdominal aorta segment; the one or more processors implementing an anterior decoder decoding the visual features to generate a plurality of anterior calcification scores, each anterior calcification score associated with an anterior portion of one abdominal aorta segment; the one or more processors implementing a posterior decoder that decodes the visual features to generate a plurality of posterior calcification scores, each posterior calcification score associated with a posterior portion of one abdominal aorta segment; each of the forward and backward decoders is a trained machine learning decoder, the forward decoder being trained separately from the backward decoder; The method comprising determining a disease risk category based on the determined plurality of anterior and posterior calcification scores.

18. 10. A method for diagnosing or predicting the prognosis of a disease or condition in a human, comprising the steps of classifying a calcification score obtained by the method for determining abdominal aortic calcification according to any one of claims 1 to 9 into a risk category, and identifying the prognosis of the disease from the determined risk category.

19. 19. The method of claim 18, wherein the disease or condition comprises CVD, diabetes, dementia, or osteoporosis, risk of falls and / or fractures in later life.

20. 10. A program for controlling one or more processors, the program comprising instructions stored on a non-volatile medium for controlling the one or more processors to implement the method of claim 1 or to operate as the system of claim 9.

Images