Method and system for measuring paraspinal muscle parameters using magnetic resonance imaging (MRI)

Abstract

This disclosure provides a method and system for measuring paraspinal muscle parameters using MRI. Two-dimensional MRI image slices of the spine of interest may be acquired. A slice selection model may identify one or more representative slices that are located in or adjacent to physiological regions corresponding to specific biomechanical environments at each vertebral level of the spine of interest. A muscle segmentation model may segment each representative slice. Segmentation of a representative slice may include generating an image segmentation mask that includes multiple pixels corresponding to the paraspinal muscles within the representative slice and excludes multiple pixels corresponding to metallic artifacts and / or biomaterials other than the paraspinal muscles within the representative slice. A measurement extraction model may extract measurements of one or more parameters of the paraspinal muscles of interest based on multiple pixels contained in the generated image segmentation mask of the representative slice.

Description

TECHNICAL FIELD

[0001] Cross-reference of related applications This application claims priority and benefit of U.S. Patent Application No. 18 / 470,269, filed on September 19, 2023 and entitled "PREDICTING PARASPINAL MUSCLE SARCOPENIA USING MAGNETIC RESONANCE IMAGING", the entire contents of which are incorporated herein by reference in their entirety.

[0002] Field of Technology The present disclosure relates generally to measuring paraspinal muscle group parameters, and more specifically to techniques for measuring a subject's paraspinal muscle group parameters using magnetic resonance imaging (MRI). BACKGROUND ART

[0003] background The health of the spinal components plays an important role in spinal function. When the spinal unit of the vertebral body and its supporting structures, the paraspinal muscle groups, ligaments, and tendons, function normally, this structural system that functions in coordination can withstand forces exceeding 1000 times the load at which individual components fail alone. Dysfunction of the components of the spinal unit can lead to increased pain, increased risk of fracture, spinal instability, and increased complications after spinal surgery.

[0004] Currently, the only standardized method for evaluating the health of spinal units is based on bone mineral assessment of the vertebral bodies. Bone density assessment using both dual-energy X-ray absorptiometry (DEXA) and quantitative computed tomography (q-CT) can estimate the mineral content within the vertebral body. At present, there is no standardized method for evaluating other soft tissue components of the spinal unit.

[0005] Recent studies have led to the use of magnetic resonance imaging (MRI) to develop methodologies for measuring paraspinal muscle parameters. Parameters such as functional cross-sectional area (FCSA) and fat infiltration (FI) have been shown to predict low back pain. Fat distribution and characteristics within the posterior paraspinal muscles further reflect changes in functional muscle units in aging and degenerative disease states. Aging and disease states affect not only the overall FI but also the fat distribution patterns within individual posterior paraspinal muscles. The posterior spinal muscles can be classified into multifidus (MF) and erector spinae (ES) groups. MFs develop both portal (medial) FI and marbled FI within the muscle. ESs accumulate marbled FI throughout the muscle. Furthermore, there are individual differences in the health of individual muscle cells within the muscle group.

[0006] Traditionally, MRI-based muscle images are pixelated, and tissue-specific threshold levels are set to distinguish between muscle, fat, and fascia. Numerical computing programs can then calculate FCSA and FI. However, conventional methods for evaluating paraspinal muscles based on MRI are prone to errors due to inconsistencies and variability in the selection of MRI slices. Furthermore, conventional methods that use pattern recognition techniques to assess fat distribution and the Sobel index are error-prone and unreliable.

[0007] The presence of metal implants in a patient negatively impacts the image quality and reliability of magnetic resonance imaging (MRI). Spinal implants are a common feature in patients with severe spinal pathology and cause a halo effect that reduces the ability to interpret pixelated images using conventional software.

[0008] Conventional segmentation techniques that distinguish between fat and muscle based on image pixelation and thresholding do not provide a quantitative geographic mapping of fat patterns and distributions within muscle with high reliability and reproducibility. Such techniques do not reliably classify muscle into distribution models.

[0009] The aforementioned examples and limitations of the related technology are intended to be illustrative and non-exclusive, and are not considered “prior art.” Other limitations of the related technology will become apparent to those skilled in the art by reading the specification and examining the drawings. [Overview of the project]

[0010] overview According to one aspect of the present disclosure, a method for measuring paraspinal muscle group parameters includes: acquiring multiple two-dimensional magnetic resonance imaging (MRI) slices of a target spine; identifying one or more representative slices from the multiple MRI slices by a slice selection model (each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine); segmenting each of the one or more representative slices by a muscle segmentation model (the segmentation of each representative slice includes generating an image segmentation mask that includes multiple pixels in each representative slice corresponding to the paraspinal muscle group and excludes multiple pixels in each representative slice corresponding to metallic artifacts and / or other biological material of the target other than the paraspinal muscle group); and extracting one or more measurements of each of the one or more parameters of the target paraspinal muscle group by a measurement extraction model 230 based on multiple pixels contained in the generated image segmentation mask of each of the one or more representative slices.

[0011] According to another aspect of the present disclosure, a system for measuring paraspinal muscle group parameters includes one or more processing units and one or more storage devices that store instructions causing the system to perform calculations when performed by the one or more processing units. The calculations include acquiring multiple two-dimensional magnetic resonance imaging (MRI) slices of the spine of interest, identifying one or more representative slices from the multiple MRI slices by a slice selection model (each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the spine of interest), segmenting each of the one or more representative slices by a muscle segmentation model (the segmentation of each representative slice includes generating an image segmentation mask that includes multiple pixels in each representative slice corresponding to the paraspinal muscles and excludes multiple pixels in each representative slice corresponding to metallic artifacts and / or other biological material of interest other than the paraspinal muscles), and extracting one or more measurements of each of the one or more parameters of the paraspinal muscles of interest by a measurement extraction model 230 based on multiple pixels included in the generated image segmentation mask of each of the one or more representative slices.

[0012] In another aspect of the present disclosure, a computer-readable storage medium is provided. This computer-readable storage medium stores instructions configured to cause a system to perform calculations when executed by one or more processing units of the system. These calculations include acquiring multiple two-dimensional magnetic resonance imaging (MRI) slices of a target spine, identifying one or more representative slices from the multiple MRI slices by a slice selection model (each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine), segmenting each of the one or more representative slices by a muscle segmentation model (the segmentation of each representative slice includes generating an image segmentation mask that includes multiple pixels in each representative slice corresponding to the paravertebral muscles and excludes multiple pixels in each representative slice corresponding to metallic artifacts and / or other biological material of the target other than the paravertebral muscles), and extracting one or more measurements of each of the one or more parameters of the target paravertebral muscles based on multiple pixels contained in the generated image segmentation mask of each of the one or more representative slices by a measurement extraction model 230.

[0013] The above and other preferred features, including various novel implementation details and combinations of events, are further described in the claims with reference to the accompanying drawings. It should be understood that the specific systems and methods described herein are illustrative and not limiting. As those skilled in the art will understand, the principles and features described herein can be applied to many different embodiments without departing from the scope of the invention. As can be understood from the above and below descriptions, each feature described herein, and any combination of two or more such features, is included in the scope of this disclosure, provided that the features included in such combination are not contradictory. In addition, any feature or combination of features may be specifically excluded from any embodiment of the invention.

[0014] The above summary is intended to assist readers in understanding this disclosure and does not limit the scope of the claims in any way. [Brief explanation of the drawing]

[0015] The drawings attached as part of this specification illustrate and teach the principles described herein, together with the general description above and the detailed description of the preferred embodiments below.

[0016] [Figure 1] This is a workflow diagram illustrating a clinical environment using a paraspinal muscle measurement system in several embodiments. [Figure 2] This is a block diagram of a paraspinal muscle measurement system according to several embodiments. [Figure 3A] This is a flowchart of a method for training a machine learning model to identify MRI slices corresponding to specific anatomical regions, according to several embodiments. [Figure 3B] This is a block diagram of a neural network model for slice selection, according to several embodiments. [Figure 4A] This is a flowchart of a method for training a machine learning model to perform muscle segmentation in MRI slices with metal subtraction, according to several embodiments. [Figure 4B] This is a block diagram of a machine learning model for muscle segmentation with metal subtraction, according to several embodiments. [Figure 5A] This is a flowchart of a method for training a machine learning model to extract paraspinal muscle parameter measurements from segmented MRI slices, according to several embodiments. [Figure 5B] This is a block diagram of a machine learning model used to extract paraspinal muscle group parameters from MRI slices, according to several embodiments. [Figure 6]is a flowchart of a method for measuring paraspinal muscle group parameters according to some embodiments. [Figure 7A] is an axial MRI image slice of a subject's spine. [Figure 7B] is the axial MRI image slice of Figure 7A with an image segmentation mask. [Figure 8] is a block diagram of an exemplary computer system according to some embodiments.

[0017] It is to be understood that for the purpose of concise and clear illustration, reference numerals may be repeated among the drawings to indicate corresponding or similar elements where considered appropriate. [Mode for Carrying Out the Invention]

[0018] Detailed explanation The present disclosure describes methods and systems for measuring paraspinal muscle group parameters using magnetic resonance imaging (MRI). Furthermore, numerous specific details are set forth to provide a thorough understanding of the examples described herein. However, those skilled in the art will appreciate that the embodiments described herein may be practiced without these specific details.

[0019] Motivation and advantages of some embodiments The present inventors have recognized and understood that the ability to accurately and reproducibly measure paraspinal muscle group parameters (e.g., cervical, thoracic, and / or lumbar paraspinal muscle group parameters) relies on selecting image slices (e.g., MRI slices) at standardized anatomical and biomechanical sites in each region of the spine. Given anatomical variations and transitional anatomical structures, conventional methodologies that apply a numbering system to each spinal unit are known to lack reliability. As a result, inter-reader variability and errors can be significant when evaluating paraspinal muscle groups.

[0020] The measurement methods described herein involve segmenting the paraspinal muscles at standard anatomical locations within the spine. Using a large patient database and studies, the inventors identified and validated key biomechanical environments that serve as indicators for assessing the health of paraspinal tissue. For example, their research has shown that both C3 and C7 are important biomechanical points for interpreting the paraspinal muscles in the cervical spine. These are important sites for determining head position and the quality of life and health of the cervical spine. Their research has also shown that the upper vertebral endplate of L4 (the second lowest motor segment) is an important biomechanical site for measuring lumbar paraspinal muscle parameters.

[0021] The presence of metal implants in a patient's body (e.g., within the spine) negatively affects the accuracy of paraspinal muscle parameter measurements. The inventors have developed a method to avoid these negative effects by reliably distinguishing between muscle and metal during the image segmentation process.

[0022] The inventors have developed a method for reliably defining paraspinal muscle quality (PMQ) based on the ratio of different spinal tissue types (e.g., fat and cerebrospinal fluid). The relationship between fat distribution pattern, particle size (Sobel index), and PMQ reflects disease state, spinal degeneration parameters, and patient demographics. The inventors' research has shown that these parameters influence spinal health and can have a significant impact on patients' health and quality of life.

[0023] This specification discloses systems and methods for reliably selecting specified biomechanical sites within the spinal region, image-segmenting major paraspinal muscle groups, and accurately measuring paraspinal muscle parameters such as functional area, fat infiltration, fat distribution, Sobel index, and PMQ. This disclosure also describes systems and methods for reliably and accurately measuring paraspinal muscle parameters even in environments with metal spinal implants. The embodiments of the methods and systems described herein are applicable not only to the evaluation of paraspinal muscles but also, more broadly, to the evaluation of muscles and tumors in other anatomical regions in native cases and in environments with metal implants.

[0024] Several embodiments of improved methods for satellite design As shown in Figure 1, the clinical environment for assessing spinal health includes a magnetic resonance imaging (MRI) scanner 110 and a paraspinal muscle measurement system 120. During operation, the patient 105 enters the MRI scanner 110, and images 115 (e.g., two-dimensional image slices) of a specific part of the patient's body (e.g., the spine) are generated. These images 115 are provided as input to the paraspinal muscle measurement system 120, which processes the images using the techniques described below to obtain reliable and accurate measurements of paraspinal muscle parameters. As used herein, “paraspinal muscle” refers to both the paraspinal muscle tissue and any fat contained within that paraspinal muscle tissue, unless otherwise indicated.

[0025] The measurement system 120 can be applied to MRI slices representing the whole or a portion of an MRI scan of the spine. For example, the measurement system 120 can be applied to one or more (e.g., all) MRI slices at the spinal level to measure paraspinal muscle parameters. In some embodiments, the measured paraspinal muscle parameters are appropriate for characterizing the degree of sarcopenia (e.g., systemic sarcopenia and / or sarcopenia in any paraspinal muscle group of interest). Examples of relevant paraspinal muscle parameters that can be measured include, but are not limited to, functional cross-sectional area, fat infiltration, and / or fat distribution measurements (e.g., marble ratio, Sobel index, etc.).

[0026] As shown in Figure 2, the paraspinal muscle measurement system 120 may include a slice selection system 210 (described below with reference to Figures 3A and 3B), a muscle segmentation system 220 (described below with reference to Figures 4A and 4B), and a measurement value extraction system 230 (described below with reference to Figures 5A and 5B). An MRI slice 115 of the patient's spine (e.g., the entire spine or one or more spinal levels) is provided as input to the slice selection system 210, and an MRI slice 215 of a specific anatomical region of the spine corresponding to a particular biomechanical environment is selected.

[0027] In some embodiments, the slice selection system 210 selects a plurality of MRI slices 215 (e.g., one slice in a specific anatomical spinal region corresponding to a specific biomechanical environment at each of two or more spinal levels, or two or more slices in or near a specific anatomical spinal region corresponding to a specific biomechanical environment at a spinal level). Examples of these anatomical regions include (i) the superior vertebral endplate of the biomechanical environment corresponding to the L4 vertebral body (at the lumbar level), (ii) the superior vertebral endplate of the biomechanical environment corresponding to the C7 vertebral body (at the cervical level), and (iii) the superior vertebral endplate of the biomechanical environment corresponding to the C3 vertebral body (at the cervical level). In the case of L4, the paraspinal muscles include the psoas major, erector spinae, and multifidus. In some embodiments, the selected MRI slice 215 is an axial image perpendicular to the posterior cortex of the vertebral body in a specific anatomical spinal region. For example, at the lumbar level, the selected MRI slice 215 may be an axial image perpendicular to the posterior cortex of the L4 superior vertebral endplate. The inventors observed that orthogonal orientation of the image helps in accurately measuring paraspinal muscle parameters expressed as absolute values ​​(e.g., FCSA). For paraspinal muscle parameters expressed as ratios or percentages (e.g., fatty infiltration), sufficiently accurate measurements can be obtained even with MRI slices that are deviated from orthogonal to the posterior cortex of the vertebral body (e.g., up to 45 degrees).

[0028] The muscle segmentation system 220 performs image segmentation of the major paraspinal muscle groups in selected MRI slices 215, thereby generating segmented MRI slices 225 corresponding to each of the selected MRI slices 215. In some embodiments, the muscle segmentation system 220 accurately performs muscle segmentation even in environments where metal spinal implants are present.

[0029] The measurement extraction system 230 processes the segmented MRI slice(s) to extract measurements 235 of paraspinal muscle parameters (e.g., functional cross-sectional area, fat infiltration, fat distribution, etc.). In some embodiments, the extracted measurements are accurate even in environments where metal spinal implants are present.

[0030] As shown in Figure 3A, a method 300 for training a machine learning model to identify two-dimensional axial MRI image slices ("MRI slices" or "slice") corresponding to a specific anatomical region may include steps 302–308. In step 302, a set of MRI slices 115 representing a three-dimensional MRI scan of the patient's spine is obtained. Each slice may be two-dimensional (e.g., height × width), in which case the third dimension is the number of slices (along the depth axis). The set of MRI slices may include slices of the entire patient's spine or slices of one or more spinal levels.

[0031] In step 304, the machine learning model selects a candidate slice from the set of slices that is predicted to be a representative slice 215 in a specified physiological region (e.g., an anatomical spine region) corresponding to a particular biomechanical environment. In some embodiments, the physiological region is selected based on its suitability for providing reliable and accurate measurements of paraspinal muscle parameters. For example, in the lumbar spine, the physiological site (e.g., anatomical site) suitable (e.g., optimal) for measuring paraspinal muscle parameters corresponds to the upper vertebral endplate of the L4 vertebra. In the cervical spine, the physiological site (e.g., optimal) for measuring paraspinal muscle parameters corresponds to the upper vertebral endplate of the biomechanical environment corresponding to the C7 and C3 vertebral bodies. In the thoracic spine, the physiological site (e.g., optimal) for measuring paraspinal muscle parameters corresponds to the central apical segment, which is a unique biomechanical environment that can occur in different anatomical sites (e.g., T5-T8 vertebrae) based on the individual's unique thoracic anatomical structure (e.g., the individual's thoracic kyphosis). In some embodiments, the optimal physiological site for measuring paraspinal muscle parameters of the thoracic spine corresponds to the point of maximum kyphosis, which is generally between T5 and T8.

[0032] In some embodiments, the set of slices provided as input to the model includes slices originating from two or more vertebral levels, and the model simultaneously selects candidate slices from each of these two or more vertebral levels. In some embodiments, the set of slices includes slices originating from two or more vertebral levels, but the model selects candidate slices from a single vertebral level. In some embodiments, the set of slices includes slices originating from a single vertebral level, and the model selects candidate slices from that vertebral level. In the latter case, additional batches of slices originating from other vertebral levels of the patient may be provided as input to the model, and the model may select candidate slices from those batches for the corresponding vertebral levels.

[0033] The inventors have observed that candidate slices located two slices above the optimal physiological site, one slice above the optimal physiological site, one slice below the optimal physiological site, or two slices below the optimal physiological site may be suitable for reliably and accurately measuring paraspinal muscle parameters, because the paraspinal muscle parameters tend to be sufficiently similar across the entire region represented by a range of ±2 slices from the optimal physiological site.

[0034] This model selects candidate slices based on features learned from representative slices and contextual information derived from other non-representative slices included in the set of slices used to train the model. In some embodiments, the model includes a first stage in which each slice is converted into a coded representation. Initially, the first stage of the model converts each two-dimensional slice into a coded representation. The coded representation can help the model learn and identify patterns, textures, and / or spatial relationships within the slice.

[0035] In some embodiments, the first stage includes a convolutional neural network (CNN), a transformer, and / or an encoder-decoder architecture (e.g., an autoencoder). The CNN may have two or more levels, where the first level transforms pixels into representations using learned convolutional weights, the next level transforms the previous representations into new transformations, and so on. The first stage of the model may perform one or more preprocessing tasks (e.g., sequence padding, data augmentation). Data augmentation operations can increase the versatility of the training dataset, thereby improving the generalization ability of the model. Examples of image data augmentation tasks include slice deletion, probabilistic modification of start and end points, image rotation, affine transformation, translation, image inversion, contrast modification, brightness modification, gamma modification, clipping of high and / or low pixel values, elastic deformation, region blurring, zoom, cropping, stretching, image sharpness modification, and noise introduction. The slice coding produced by the first stage of the model represents a sequence with consistent spatial relationships between slices.

[0036] The series of slice-coding data are then provided as input to a second stage of the model. In some embodiments, the second stage of the model includes an iterative recurrent neural network (iterative RNN) that uses the spatial context from the order of the slice codings to determine the probability that each slice coding is located in a specified physiological site. In this second stage of the model, the slice corresponding to the coding that has the highest probability of being in the specified physiological site is selected as a candidate slice.

[0037] In some embodiments, candidate slices may be selected in the second step of this model using a transformer or other preferred type of model instead of an iterative RNN. In some embodiments, candidate slices may be selected in the second step of this model using additional features (e.g., coding). Such features may include contextual features that can influence the location or appearance of the physiological site of interest (e.g., features indicating contextual information such as the patient's height, weight, sex, age, and health status).

[0038] We have described an example of a slice selection model having a two-stage architecture in which coding is generated in the first stage and candidate slices are selected based on the coding in the second stage. In some embodiments, the process of obtaining the coding of slices and selecting candidate slices using a transformer structure may be integrated into an architecture such as a Vision / Visual Transformer (ViT) which includes feature extraction, linear embedding, transformer layers, and a classification head for coding.

[0039] An alternative technique for medical slice selection quantifies the probability that a slice is a target slice (e.g., the slice located at or closest to a specified physiological site) by applying a one-to-one slice-probability model to each slice in a three-dimensional image and selecting the slice with the highest probability as a candidate slice. Slice selection method 300 improves upon this alternative technique by using a clinically informed many-to-one slice-versus-probability approach, where the evaluation of each slice is based not only on the individual slice but also on other slices. This approach uses two or more slices (e.g., all slices) to predict which single slice is the target slice. In the many-to-one slice-versus-probability approach, a series of numerous slices may be compared and candidate slices may be selected using context derived from multiple (e.g., all) slices in the three-dimensional image.

[0040] In step 306, the candidate slice is compared to the actual representative slice 307 ("ground truth slice"). In some embodiments, comparing the candidate slice to the ground truth slice includes measuring the number of slices between the candidate slice and the ground truth slice in the sequence of slices.

[0041] In step 308, the model parameters may be adjusted based on a comparison between the candidate slice and the actual representative slice 307. In some embodiments, backpropagation with cross-entropy loss and stochastic optimization is used to adjust the internal weights of the model's neural network during training. In some embodiments, the loss is scaled to be proportional to the difference between the selected candidate slice and the actual representative slice ("ground truth" slice). This difference is calculated as the number of slices from the actual representative slice to the selected candidate slice and may be signed or unsigned.

[0042] A many-to-one slice-versus-probability approach to slice selection models can be used to perform tasks other than selecting the MRI slice closest to a specified physiological site. For example, some embodiments of slice selection models may be configured to select one or more slices corresponding to a particular biomechanical environment determined through image features (e.g., by a medical professional). Some embodiments of slice selection models may be configured to select one or more slices corresponding to an anatomical correlation of a pathological finding (but not necessarily corresponding to an image feature) as determined by a medical professional. Such pathologies may include, but are not limited to, primary tumors, metastases, fractures, connective tissue changes, fibrous changes (e.g., cirrhosis), and fluid accumulations (e.g., cysts, hematomas, pleural effusion, pericardial effusion, ascites). Some embodiments of slice selection models may be configured to select one or more slices corresponding to an anatomical site suitable for downstream tasks such as segmentation, sizing, and identification of pathological sites. An example of such a downstream task is to characterize the degree of sarcopenia using an MRI slice corresponding to the upper vertebral endplate in a biomechanical environment representing the L4 vertebral body. Other examples of anatomical sites include C3 for cranial nerve parameters and C7 for balance measurements.

[0043] Examples of slice selection models having a two-stage architecture or a ViT architecture are described. In some embodiments, the slice selection model has a Siamese neural network architecture. In such embodiments, slices in a three-dimensional medical image (e.g., an MRI image) are provided as input to a Siamese neural network, which calculates the probability that each slice is a target slice (e.g., a slice located in or near a specified physiological site) and selects the slice with the highest probability as the target slice. This model calculates the probability that each candidate slice C is a target slice based on its similarity to one or more actual representative slices ("ground truth" slices) located in or near a specified physiological site. Thus, the Siamese neural network model selects candidate slices by comparing each slice to an actual representative slice(s) and using the context derived from multiple (e.g., all) slices in the three-dimensional image to select the candidate slice that is most likely to be the target slice.

[0044] In some embodiments, the Siamese neural network model is supported by two convolutional neural networks (CNNs) that share identical or very similar parameters and operate synchronously to process different image inputs in parallel. These CNNs are structured to analyze and extract hierarchically complex features from input images through a series of convolutional, pooling, and fully connected layers, ultimately yielding a feature vector representation of the analyzed image. During the training phase, two image pairs—a representative target image (ground truth, spine level) and a non-representative image—are propagated through the network, and a feature vector is computed for each image. A loss function (e.g., control loss) is then used to calculate the dissimilarity between the derived feature vectors, incentivizing the network to minimize the distance between similar spine level instances and maximize the distance between different spine level instances. As a result, through iterative training cycles, the network improves its ability to identify and classify specific spine levels shown in a given MRI slice with increasing accuracy.

[0045] Other embodiments include alternative loss functions such as cross-entropy loss and triplet loss. The cross-entropy loss function serves to minimize the discrepancy between the true vertebral level classification and the predicted vertebral level classification, while the triplet loss function works by not only distinguishing between positive and negative examples but also ensuring that positive examples are closer to the anchor slice by a certain margin than negative examples.

[0046] In some embodiments, the slice selection module is configured to select one or more slices corresponding to pathologies in local anatomical sites determined by a medical professional. Such pathologies include, but are not limited to, tumors, masses, fractures, connective tissue changes, and fluid accumulations. In some embodiments, the slice selection module is configured to select one or more slices corresponding to quantitative imaging biomarkers (QIBs) associated with vertebral neighborhood structures (e.g., aortic calcification, carotid artery stenosis). For subjects having such vertebral neighborhood structures, such QIBs may be identified based on analysis of medical images obtained using spinal MRI and conventional imaging techniques suitable for interpreting calcium (e.g., CT angiography, CT, X-ray, etc.).

[0047] Figure 3B shows an example of a trained slice selection model 350. As described above, the slice selection model 350 may include a first stage 352 and a second stage 356. The first stage 352 may include a convolutional neural network (CNN) that generates codings 354 for a set of MRI slices 115 of the patient's spine. The second stage 356 may include a recurrent neural network (RNN) 356 configured to select a candidate slice 358 that has the highest probability of being the target slice, based on the codings of all slices in the input set. The target slice may be a slice of a target physiological site in the body corresponding to a particular biomechanical environment or a slice in its vicinity (e.g., within ±2 slices). In some embodiments, the trained selection model 350 is a convolutional recurrent neural network (CRNN).

[0048] In some embodiments, the CNN in the first stage 352 of the slice selection model 350 is a two-dimensional CNN configured to encode MRI slices. The performance of the CNN may be improved by using group normalization (or batch normalization), dropout, nonlinear activation functions, and skip (or residual) connections.

[0049] In some embodiments, the CNN and / or RNN may be initialized using pre-trained model weights during the training process. Pre-trained CNN weights may be provided by models including, but not limited to, VGG, Inception, ResNet, MobileNet, or EfficientNet. Pre-trained CNN weights may be generated using a pre-training step, or provided by a Siamose neural network or a contrast learning process.

[0050] In some embodiments, the RNN may have an iterative network architecture. For example, the recurrent weights may use a GRU architecture or an LSTM architecture. In some cases, the iterative network architecture is based on a transformer network. In some cases, the iterative network is multilayered and bidirectional. In some embodiments, the final output sequence of the RNN is passed through a SoftMax activation function (or a Sigmoid activation function in the case of a binary class), a binary cross-entropy loss is applied to each output, and the loss is minimized by ADAM optimization with backpropagation and gradient clipping.

[0051] In some embodiments, the trained slice selection model exhibits accuracy of over 85%, 85% to 90%, over 90%, 90% to 95%, or over 95% when selecting slices from the target physiological site or its vicinity (e.g., within ±2 slices).

[0052] Figure 4A shows a method 400 for training a machine learning model to perform muscle segmentation in an MRI slice with metal subtraction. The trained model may be configured to segment and separate the paraspinal muscles in a representative MRI slice 215 (e.g., a slice selected by a slice selection model 350). In this context, “metal subtraction” refers to the process of segmenting around metal in an MRI slice, so that the metal is not included in the portion of the slice subsequently analyzed for measuring muscle parameters. This technique can be effectively applied to spinal MRI images of subjects with any suitable metal implants, e.g., titanium pedicle screws, intervertebral cages, spinal cord stimulators, etc.

[0053] Training method 400 may include steps 402-408.

[0054] In step 402, a representative MRI slice 215 of the target spine level is obtained.

[0055] In step 404, the machine learning model 450 (hereinafter referred to as the “Segmentation Model”) performs metal subtraction and muscle segmentation on a representative MRI slice 215. In some embodiments, the Segmentation Model 450 is a U-net based CNN model. In some embodiments, the U-net based CNN model includes transformation-based modifications and varying degrees of connectivity between levels. In some embodiments, the U-net based CNN model uses group normalization, skip connections, and attention layers. In some embodiments, the U-net based CNN model operates similarly to a V-net, replacing upsampling and downsampling pooling layers with convolutional layers. In some embodiments, the U-net based CNN model includes nested and skip connections for connecting encoders and decoders via a series of nested, tightly coupled convolutional blocks. In some embodiments, the U-net based CNN model incorporates recurrent convolutional layers. In some embodiments, an attention gate is added to the U-net based CNN model, which guides the model to focus on more relevant regions and feature selectivity. In embodiments including transformers, the transformers may include, but are not limited to, vision transformers and / or shifted window (swin) transformers. In some embodiments, the U-net-based CNN model includes full-scale skip connections and classification guidance modules.

[0056] The metal subtraction feature of this model removes image artifacts associated with metal implants (e.g., plates, rods, and screws used in spinal fusion), so that pixels corresponding to or distorted by metal are not analyzed during subsequent determination of paraspinal muscle parameters. Since extracting accurate muscle parameter measurements from metal-distorted images is extremely difficult, metal subtraction is crucial for obtaining accurate paraspinal parameter measurements for many subjects with metal implants. While machine learning models can be trained to extract accurate muscle measurements from slices containing metal distortion, such training relies on the availability of a training dataset containing a suitable number of MRI slices with metal artifacts. We have noted that such datasets are generally small in size and poorly labeled, making it difficult to train models to extract accurate muscle measurements from slices containing metal distortion using available training data.

[0057] In contrast, training the segmentation model 450 to perform metal subtraction is more practical because suitable training datasets are more readily available. We have observed that the segmentation model 450 can be trained to reliably and accurately perform metal subtraction on MRI slices using synthetic training data (e.g., MRI images artificially modified to simulate the presence of metal artifacts) and / or metal training data (e.g., MRI images in which metal artifacts are present).

[0058] Conventional methods for processing medical images containing artifacts caused by medical devices involve individually training models on datasets of patients who consistently exhibit medical device artifacts. This disclosure presents two approaches that improve the efficiency and accuracy of conventional methods by enabling the incorporation of data from patients who do not wear medical devices. The first approach utilizes digital artifact correction, which corrects images of patients without medical devices by adding the device to anatomical sites corresponding to the placement of the device, and modifies the corresponding ground truth segmentation mask so that it does not include the boundaries set by the digital artifact. The appearance and location of digital artifacts can be determined by medical professionals with experience in placing such devices, or by utilizing deep learning techniques such as generative adversarial networks (GANs).

[0059] The second approach utilizes transfer learning. In this approach, the dataset is divided into patients with and without medical devices. Each dataset is manually segmented by a medical professional, and the device dataset is segmented so that the devices themselves are not included. Since medical datasets typically contain many images of patients without medical devices, the non-medical device dataset is used to train an accurate segmentation model conforming to the training method in Figure 4A. The parameters of this initial model are used as the initial parameters for a second training using only the device dataset. This training iteration is consistent with the training method in Figure 4A and can be enhanced by using fixed weights, a low learning rate, and image augmentation.

[0060] Returning to step 404 of segmentation method 400, the muscle segmentation function of the model includes separating and identifying (e.g., highlighting) the paraspinal muscle group region containing fat. After performing metal subtraction and muscle segmentation, segmentation model 450 outputs a segmented map that identifies (e.g., highlights) the paraspinal muscle group.

[0061] In step 406, the segmentation map generated by the model is compared to the ground truth segmentation map 420. The ground truth segmentation map 420 can be created manually by pixelating muscle, fat, and fascia and by tissue-specific intensity threshold levels.

[0062] In step 408, the parameters of the segmented model 450 may be adjusted based on a comparison between the model's segmentation map and the ground truth segmentation map 420. In some embodiments, a combined DICE and binary cross-entropy loss function is used to adjust the intrinsic parameter values ​​of the segmented model (e.g., neural network weights) during training. Other loss functions include individual DICE losses, individual BCE losses, boundary losses, Tversky losses, IoU / Jaccard losses, and weighted loss functions. The model weights are then appropriately updated by backpropagation using stochastic optimization.

[0063] Referring to Figure 4B, an example of the segmentation model 450 is shown. In the example in Figure 4B, the segmentation model 450 is a U-net based CNN model consisting of multiple convolutional layers 452, followed by downsampling with a pooling layer 454, and then another set of convolutional layers 456, and so on. Each sequence of convolutional layers learns more abstract features of the paraspinal muscles in an MRI slice. In some embodiments, downsampling may consist of learned convolutional weights, mean pooling, or max pooling. Once this downsampling is complete, these latent representations 458 are upsampled 460 to create a segmentation map 480 of the paraspinal muscles. This upsampling is performed with several convolutional layers 462, followed by an upsampling layer 464, followed by another set of convolutional layers 466, and so on. The output of the final upsampling layer is a map of the probability that each pixel in the image is part of a particular paraspinal muscle group, and the model selects the class with the highest probability for each pixel. The class assigned to a pixel may indicate whether that pixel is part of a muscle, and if so, the assigned class may indicate the type of muscle.

[0064] A segmentation model 450 may be trained to accurately perform a variety of tasks, including but not limited to the automated segmentation of subcutaneous fat / adipose tissue, visceral fat / adipose tissue, whole-body adipose tissue, skeletal muscle, smooth muscle, cardiac muscle, and / or bone, using the training method 400 shown in Figure 4A.

[0065] In some embodiments, data augmentation may be performed on the training data (e.g., representative slices) before training the segmentation model 450. Data augmentation operations can increase the versatility of the training dataset and thereby improve the generalization ability of the segmentation model. Examples of image data augmentation tasks include slice deletion, probabilistic modification of start and end points, image rotation, affine transformation, translation, image inversion, contrast modification, brightness modification, gamma modification, clipping of high and / or low pixel values, elastic deformation, region blurring, zoom, cropping, stretching, image sharpness modification, and noise introduction.

[0066] In some embodiments, the segmentation model 450 may include an encoder that codes representative image slices. The encoder may be, for example, a two-dimensional convolutional neural network (CNN). The performance of the CNN may be improved by group normalization (or batch normalization), dropout, nonlinear activation functions, and skip (or residual) connections.

[0067] In some embodiments, the second round of learning in the transfer learning process is effective by using fixed weights and a low learning rate.

[0068] In some embodiments, the loss function for training the segmentation model is a combination of DICE loss and binary cross-entropy. Other potential loss functions in semantic segmentation include individual DICE losses, individual BCE losses, boundary losses, Tversky losses, IoU / Jaccard losses, and weighted loss functions. In some embodiments, the loss is minimized by ADAM optimization with backpropagation and gradient clipping.

[0069] The inventors have collected experimental data showing that several embodiments of the muscle segmentation model 450 reduce measurement errors of various paraspinal muscle group parameters by 3% to 14% in subjects where metal artifacts are observed in spinal MRI slices.

[0070] Referring to Figure 5A, a method 500 for training a machine learning model to extract paraspinal muscle parameter measurements from segmented MRI slices of a subject is shown. The trained model may be configured to output measurements of paraspinal muscle parameters (such as functional cross-sectional area, fat infiltration, and fat distribution). In some embodiments, the trained model is configured to output sarcopenia features in the subject.

[0071] Training method 500 may include steps 502-508.

[0072] In step 502, segmented representative MRI slices 480 for the level of the target spine and actual representative slices 307 ("ground truth slices") derived from MRI slices 115 of the original set of targets are obtained. The segmented representative MRI slices 480 include a segmentation mask.

[0073] In step 504, a machine learning model 550 (referred to herein as the “measurement extraction model”) extracts measurements of paraspinal muscle parameters (e.g., functional cross-sectional area, fat infiltration, fat distribution, etc.) from segmented representative MRI slices 480.

[0074] To extract these measurements, the measurement extraction model 550 applies the segmentation mask generated by the segmentation model 450 to a representative MRI slice 307, thereby including all pixels within the segmentation mask (e.g., retaining their values ​​unchanged) and excluding all pixels outside the segmentation mask (e.g., screening them out, marking them as non-contributing, etc.). This step generates a superimposed mask image, which passes through the first stage of the model (e.g., a multilayer convolutional neural network with trained weights and kernels) to transform the superimposed mask image into a representation in the latent feature space. This abstract intermediate representation is then reconstructed into linear vectors and used as input to the second stage of the model (e.g., a deep neural network (DNN) with trained weights). This DNN may include an input layer, one or more hidden linear layers, and one or more (e.g., three) output layers. In some embodiments, the DNN's performance is improved by adding dropout, ReLU activation, and batch normalization. The second stage of the model may involve extracting 580 measurements of paraspinal muscle parameters (e.g., functional cross-sectional area, fat infiltration, and fat distribution) from the coding generated in the first stage of the model.

[0075] In some embodiments, the measurement extraction model 550 may use a bias field correction algorithm such as N4 or MICO to remove low-frequency intensity fluctuations or inaccurate intensity gradients in the MRI images.

[0076] An example is described in which measurements of paraspinal muscle parameters are extracted from a segmented representative MRI slice 480 using the measurement extraction model 550. Other techniques may be used to measure paraspinal muscle parameters. In some embodiments, the extraction of measurements of specific paraspinal muscle parameters, such as fat infiltration, involves "thresholding" (classifying pixels as "muscle" or "fat" based on their pixel values ​​within the segments of the MRI slice 480 designated as paraspinal muscles by the image segmentation mask), and then dividing the number of pixels in the segments classified as fat by the total number of pixels (sum of fat pixels and muscle pixels). This numerical calculation method may also be applied to other parameters.

[0077] More formally, "thresholding" is the process of converting a grayscale image (or part thereof) into a binary image. Thresholding involves setting a threshold that helps convert a grayscale image into a binary image, classifying pixel intensities higher and lower than this value into separate categories (e.g., paravertebral muscles and fat). Examples of thresholding methods include, but are not limited to, the Otsu method, the Isodata (intermeans) algorithm, mean of firsts, mean of firsts + standard deviation, adaptive thresholding, Li, Triangle, Yen, mean, minimum, Niblack, and Sauvola.

[0078] The distribution of fat can also be characterized by the extent to which fat belongs to continuous large fat segments or discontinuous small fat segments. In some embodiments, this feature determination can be made by techniques such as contraction and expansion, morphological closure, depth-first search, breadth-first search, thinning, or granular analysis, but is not limited to these. In some embodiments, this feature determination can be made by unsupervised learning methods including, but not limited to, conditional random fields, clustering, local binary patterns, Gabor filters, or fractal analysis.

[0079] In some embodiments, a deep learning system configured to use thresholding techniques (e.g., the thresholding techniques described above) may be used to autonomously extract paraspinal muscle parameter measurements from a representative MRI slice 480. In some embodiments, these measurements are used for clinical purposes. Alternatively, such measurements and corresponding MRI slices can be used to train a measurement extraction model 550. When used to train the measurement extraction model 550, the deep learning system may repeatedly train the measurement extraction model 550 on batches of data by repeatedly using the data generation module and the model training module.

[0080] In step 506, the muscle parameter measurements extracted by the measurement extraction model 550 are compared with ground truth measurements 520. Ground truth measurements 520 may be obtained using conventional techniques (e.g., expert or computer-based analysis of manually segmented MRI images).

[0081] In step 508, the parameters of the measurement extraction model 550 may be adjusted based on a comparison of the measurements 580 extracted from the model with the ground truth measurements 520. The training process involves updating the model parameters to minimize a loss function that represents the difference between the extracted measurements 580 and the ground truth values ​​520. In some embodiments, the loss function(s) used to adjust the intrinsic parameter values ​​of the model may include, but are not limited to, cross-entropy, mean squared error, mean absolute error, log-likelihood, hinge loss, Huber loss, etc. The model weights are then updated by backpropagation using stochastic optimization.

[0082] Referring to Figure 5B, an example of the measurement extraction model 550 is shown. In the example in Figure 5B, the measurement extraction model 550 includes a first stage 552 and a second stage 554. In some embodiments, the first stage 552 includes a multilayer convolutional neural network (CNN) with trained weights and kernels, which transforms the superimposed mask images into representations in the latent feature space. The CNN then reconstructs this abstract intermediate representation into linear vectors, which it provides as input to the second stage 554. In some embodiments, the second stage 554 includes a deep neural network (DNN) with trained weights. This DNN may include an input layer, one or more hidden linear layers, and one or more (e.g., three) output layers. The second stage 554 of the model may extract measurements 580 of the paraspinal muscle group parameters.

[0083] Using the training method 500 in Figure 5A, the measurement extraction model 550 may be trained to accurately perform a variety of tasks, including, but not limited to, (1) detecting and evaluating sarcopenia and systemic wasting disorders (e.g., cachexia) associated with spinal degenerative changes (e.g., spinal stenosis, facet joint degeneration, instability, deformity, etc.), and (2) evaluating the quality of the paraspinal muscles. (The inventors have developed a novel muscle tissue quality assessment tool that derives parameters such as paraspinal muscle quality (PMQ), which evaluates the health of the muscle cells themselves based on the ratio of soft tissue type to cerebrospinal fluid and fat.) Some embodiments of the muscle measurement extraction model 550 may also be used to enhance the prediction of cardiovascular metabolic health markers or the prediction of mortality due to the aforementioned segmented pathologies.

[0084] Referring to Figure 6, several embodiments of Method 600 for measuring paraspinal muscle parameters are shown. Method 600 may include steps 610-640.

[0085] In step 610, a two-dimensional magnetic resonance (MRI) image slice 115 of the target spine may be acquired. The MRI image slice 115 may be acquired using any preferred technique, for example, by scanning the target spine with an MRI scanner to generate an MRI image slice, or by reading an MRI image slice stored from a local or remote computer-readable storage device.

[0086] In step 620, the slice selection model may identify one or more representative slices 215 within the MRI slice 115. Each representative slice may be located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the spine in question.

[0087] To identify representative slices 215, any suitable slice selection model may be used, including but not limited to embodiments of the slice selection model 210 described above. In some embodiments, the first-stage slice selection model generates coding for each of the MRI image slices 115. In some embodiments, the MRI image slices 115 include one or more sets of MRI image slices corresponding to each vertebral level of one or more vertebrae of interest. In some embodiments, from each set of MRI image slices, the second-stage slice detection model selects a representative slice that is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level. The selection of a representative slice for a particular vertebral level may be based on the coding of all image slices in the set corresponding to that vertebral level, or on the coding of all MRI image slices 115. Relying on the coding of multiple image slices to identify a representative MRI image slice may be referred to herein as a “many-to-one slice-to-probability approach”. Optionally, the selection of a representative slice for a given vertebral level may also be based on the order of the image slices within multiple image slices, since the order of the image slices conveys information about the position of each slice relative to the other slices (and to the vertebra of the subject).

[0088] In some embodiments, one or more representative slices include a representative slice located at or adjacent to the biomechanical upper vertebral endplate corresponding to the C7 vertebral body or the biomechanical upper vertebral endplate corresponding to the mid-C3 vertebral body at the cervical level of the spine in question. In some embodiments, one or more representative slices include a representative slice located at or adjacent to the apical segment at the thoracic level of the spine in question. In some embodiments, the apical segment of the spine is located at or adjacent to the T8, T7, T6, or T5 vertebral body. In some embodiments, one or more representative slices include a representative slice located at or adjacent to the biomechanical upper vertebral endplate corresponding to the L4 vertebral body at the lumbar level of the spine in question.

[0089] In step 630, the muscle segmentation model may segment each representative slice. Segmenting a representative slice may involve generating an image mask that includes pixels corresponding to the paraspinal muscles within the representative slice and excludes pixels corresponding to metallic artifacts and / or other biomaterials within the representative slice.

[0090] To segment the representative slice, any suitable muscle segmentation model may be used, including but not limited to embodiments of the muscle segmentation model 220 described above. In some embodiments, the muscle segmentation model is trained to exclude pixels corresponding to metal artifacts using a training dataset that includes one or more slices to which metal artifacts have been synthetically added. In some embodiments, at least a portion of the representative slice 215 contains metal artifacts (for example, because a metal implant or other device was present in the spine or another part of the body at the time the MRI image slice 115 was generated). In some embodiments, the image mask generated for a given representative slice 215 separates pixels corresponding to paraspinal muscles (e.g., all pixels corresponding to paraspinal muscles, or 99%, 98%, 95%, 90%, 85%, or more than 80% of pixels corresponding to paraspinal muscles) from pixels corresponding to metallic artifacts and pixels corresponding to biomaterials other than paraspinal muscles (e.g., all pixels corresponding to metallic artifacts and / or biomaterials other than paraspinal muscles, or 99%, 98%, 95%, 90%, 85%, or more than 80% of pixels corresponding to metallic artifacts and / or biomaterials other than paraspinal muscles).

[0091] In step 640, the measurement extraction model 230 may extract measurements of one or more parameters of the paraspinal muscle group in question based on the pixels contained in the generated image mask of each representative slice. In some embodiments, one or more parameters may include functional cross-sectional area (FCSA), fat infiltration (FI), fat distribution, Sobel index, and / or paraspinal muscle quality (PMQ).

[0092] In some embodiments, Method 600 further includes a step of characterizing the degree of sarcopenia in the subject based on the extracted measurements. In some embodiments, the measurement parameters include FCSA and FI, and Method 600 further includes a step of predicting the subject's low back pain based on the FCSA and FI measurements.

[0093] In some embodiments, the measurement parameters include fat distribution and FI within the posterior paraspinal muscles of the subject, and the method further includes assessing the state of at least one degenerative disease in the subject based on the extracted measurements of fat distribution and FI. In some embodiments, the at least one degenerative disease includes spinal stenosis, facet joint degeneration, spinal instability, or spinal deformity.

[0094] In some embodiments, the measurement parameters include the FI of the lumbar multifidus (MF) group and the erector spinae (ES) group of the subject, and the method further includes the step of predicting the outcome of lumbar fusion surgery for the subject based on the extracted measurements of the FI of the lumbar MF group and the FI of the ES group. In some embodiments, an adverse effect on the outcome of lumbar fusion surgery is predicted if the extracted measurements of FI in the lumbar MF group are at least 60% and the extracted measurements of FI in the ES group are at least 40%.

[0095] Several non-limiting examples of clinical use for measuring paraspinal muscle parameters are described in the literature listed below, which is incorporated herein by reference. The inventors recognize and appreciate that the techniques described herein can be used for measuring paraspinal muscle parameters in these and other clinical applications.

[0096] M. Muellner, A. Hughes, et al. Changes of the posterior paraspinal and psoas muscle in patients with low back pain: a 3-year longitudinal study. Eur Spine J.2023 Jun 28. doi:10.1007 / s00586-023-07814-1. Online ahead of print. PMID:37378708.

[0097] L. Schonnagel, A. Hughes, et al. Abdominal aortic calcification is independently associated with increased atrophy and fatty infiltration of the lumbar paraspinal muscles: a retrospective cross-sectional study. Eur Spine J.2023 Jun 5. doi:10.1007 / s00586-023-07783-5. Online ahead of print. PMID:37273032.

[0098] H.Haffer,A.Hughes,et al. Changes in psoas and posterior paraspinal muscle morphology after standalone lateral lumbar interbody fusion:a quantitative MRI-based analysis. Eur Spine J.2023 May;32(5):1704-1713. doi:10.1007 / s00586-023-07579-7. Epub 2023 Mar 8. PMID:36884111.

[0099] M.Muellner,A.Hughes,et al. The association between paraspinal muscle parameters and vertebral pedicle microstructure in patients undergoing lumbar fusion surgery. Int Orthop.2023 Apr;47(4):1051-1060. doi:10.1007 / s00264-022-05659-9. Epub 2022 Dec 23. PMID:36562815.

[0100] M.Moser,A.Hughes,et al. Differences in lumbar paraspinal muscle morphology in patients with sagittal malalignment undergoing posterior lumbar fusion surgery. Eur Spine J.2022 Nov;31(11):3109-3118. doi:10.1007 / s00586-022-07351-3. Epub 2022 Aug 29. PMID:36038784.

[0101] M.Muellner,A.Hughes,et al. The effect of age on psoas and paraspinal muscle morphology in patients undergoing posterior lumbar fusion surgery. Eur Spine J.2022 Oct;31(10):2619-2628. doi:10.1007 / s00586-022-07346-0. Epub 2022 Aug 19. PMID:35984509.

[0102] M.Muellner,A.Hughes,et al. Paraspinal musculature impairment is associated with spinopelvic and spinal malalignment in patients undergoing lumbar fusion surgery. Spine J.2022 Dec;22(12):2006-2016. doi:10.1016 / j.spinee.2022.07.103. Epub 2022 Aug 6. PMID:35944826.

[0103] E.Chiapparelli,A.Hughes,et al. The association between lumbar paraspinal muscle functional cross-sectional area on MRI and regional volumetric bone mineral density measured by quantitative computed tomography. Osteoporos Int.2022 Dec;33(12):2537-2545. doi:10.1007 / s00198-022-06430-x. Epub 2022 Aug 6. PMID:35933479.

[0104] M.Moser,A.Hughes,et al. The predictive value of psoas and paraspinal muscle parameters measured on MRI for severe cage subsidence after standalone lateral lumbar interbody fusion. Spine J.2023 Jan;23(1):42-53. doi:10.1016 / j.spinee.2022.03.009. Epub 2022 Mar 26. PMID:35351664. [Examples]

[0105] Data and Examples Improved accuracy through slice selection As shown in Table 1 below, the inventors collected experimental data demonstrating the extent to which several embodiments of the slice selection technique described herein can improve the accuracy of paraspinal muscle group parameters. In Table 1, the second column shows the median total cross-sectional area (CSA) of three different muscle groups (psoas major, ES, and MF) measured using MRI image slices of the L1, L2, L3, L4, and L5 vertebral bodies or their vicinity. As shown in the third column, in all cases, the measurements extracted from the L4 slice were the most accurate. For the total CSA of the psoas major, the measurements extracted from the L1 slice were of comparable accuracy, while the measurements extracted from the L2, L3, and L5 slices exhibited relative errors ranging from 15% to 61%. For the total CSA of the ES, the measurements extracted from the L1, L2, L3, and L5 slices exhibited relative errors ranging from 4% to 27%. For the total CSA of MF, measurements extracted from L1, L2, L3, and L5 slices exhibit a relative error ranging from 27% to 70%.

[0106] In Table 1, column 4 shows the median functional cross-sectional area (CSA) of three different muscle groups (psoas major, esophageal sphincter, and mid-muscle) measured using MRI image slices of the L1, L2, L3, L4, and L5 vertebral bodies or their vicinity. As shown in column 5, in all cases, the measurements extracted from the L4 slice were the most accurate. For the functional CSA of the psoas major, measurements extracted from the L1 slice showed comparable accuracy, while measurements extracted from the L2, L3, and L5 slices exhibited relative errors of 16% to 63%. For the functional CSA of the esophageal sphincter, measurements extracted from the L1, L2, L3, and L5 slices exhibited relative errors ranging from 7% to 42%. For the functional CSA of the mid-muscle, measurements extracted from the L1, L2, L3, and L5 slices exhibited relative errors ranging from 18% to 58%.

[0107] In Table 1, column 6 shows the median fat infiltration (MI) of three different muscle groups (psoas major, ES, and MF) measured using MRI image slices of the L1, L2, L3, L4, and L5 vertebral bodies or their vicinity. As shown in column 7, in all cases, the measurements extracted from the L4 slice were the most accurate. For psoas major FI, measurements extracted from the L1 slice showed comparable accuracy, while measurements extracted from the L2, L3, and L5 slices exhibited relative errors of 11% to 47%. For ES FI, measurements extracted from the L1, L2, L3, and L5 slices exhibited relative errors ranging from 9% to 31%. For MF FI, measurements extracted from the L1, L2, L3, and L5 slices exhibited relative errors ranging from 13% to 27%.

[0108] [Table 1]

[0109] Improved accuracy through metal subtraction Referring to Figure 7A, an axial MRI image slice 700 of the subject's spine is shown. In the embodiment of Figure 7A, the subject has a metal spinal implant (782, 784). As can be seen, the pixels surrounding the metal implant in MRI slice 700 are much brighter than the adjacent tissue. As previously mentioned, this "halo effect" can lead to errors in the measurement of paraspinal muscle parameters if pixels affected by the metal implant are mistakenly treated as pixels representing the paraspinal muscles.

[0110] Figure 7B shows an axial MRI image slice 700 from Figure 7A, including a segmentation mask 750 generated using one embodiment of the muscle segmentation model 450 described above. The segmentation mask 750 correctly distinguishes the psoas major muscle (752, 754), erector spinae muscles (762, 764), and multifidus muscle (772, 774). Furthermore, the segmentation mask 750 correctly excludes not only metallic spinal implants (712, 714) but also discolored pixels (782, 784) caused by the halo effect generated by the spinal implants. Therefore, the segmentation mask 750 is suitable for extracting accurate measurements of the paraspinal muscle group parameters of interest, even in the presence of metallic spinal implants.

[0111] As shown in Table 2 below, the inventors collected experimental data demonstrating the extent to which several embodiments of the muscle segmentation model 450 can reduce measurement errors for various paraspinal muscle group parameters in subjects exhibiting metal artifacts in spinal MRI slices. In Table 2, the second column shows the median and interquartile range (IQR) values ​​of total cross-sectional area (CSA) for three different muscle groups (psoas major, ES, and MF) measured using MRI image slices selected by an embodiment of the slice selection model 350. As shown in the third column, using the muscle segmentation model 450 reduces the error of the extracted measurements by 5–11%. In the embodiments of Table 2, the muscle segmentation model 450 reduces the measurement error by 5.1%, 10.5%, and 11.1% for the total CSA of the psoas major, ES, and MF, respectively.

[0112] In Table 2, column 4 shows the median and IQR values ​​of functional CSA for three different muscle groups (psoas major, ES, and MF) measured using MRI image slices selected by an embodiment of slice selection model 350. As shown in column 5, using muscle segmentation model 450 reduces the error of the extracted measurements by 5–14%. Specifically, in the embodiment in Table 2, muscle segmentation model 450 reduces the measurement error for functional CSA of the psoas major, ES, and MF by 5.2%, 10.4%, and 14.5%, respectively.

[0113] In Table 2, column 6 shows the median and IQR values ​​of fat infiltration (FI) for three different muscle groups (psoas major, ES, and MF) measured using MRI image slices selected by an embodiment of slice selection model 350. As shown in column 7, using muscle segmentation model 450 reduces the error of the extracted measurements by 3–11%. Specifically, in the embodiment in Table 2, muscle segmentation model 450 reduces the measurement error by 10.7%, 8.5%, and 2.7% for the FI of the psoas major, ES, and MF, respectively.

[0114] [Table 2]

[0115] Additional Embodiments Figure 8 is a block diagram of a computer system 800 that may be used when implementing the techniques described herein. General-purpose computers, network devices, mobile devices, or other electronic systems may also include at least a portion of system 800. System 800 includes a processor 810, memory 820, storage device 830, and input / output device 840. Each of components 810, 820, 830, and 840 may be interconnected using, for example, a system bus 850. The processor 810 can process instructions for execution within system 800. In some embodiments, the processor 810 is a single-threaded processor. In some embodiments, the processor 810 is a multi-threaded processor. In some embodiments, the processor 810 is a programmable (or reprogrammable) general-purpose microprocessor or microcontroller. The processor 810 can process instructions stored in memory 820 or storage device 830.

[0116] Memory 820 stores information within the system 800. In some implementations, memory 820 is a non-temporary computer-readable medium. In some implementations, memory 820 is a volatile memory unit. In some implementations, memory 820 is a non-volatile memory unit.

[0117] The storage device 830 can provide the system 800 with a large storage capacity. In some embodiments, the storage device 830 is a non-temporary computer-readable medium. In various different implementations, the storage device 830 may include, for example, a hard disk drive, an optical disk drive, a solid-state drive, a flash drive, or some other mass storage device. For example, the storage device may store long-term data (e.g., database data, file system data, etc.). The input / output device 840 provides input / output operations to the system 800. In some implementations, the input / output device 840 may include one or more network interface devices (e.g., Ethernet cards), serial communication devices (e.g., RS-232 ports), and / or wireless interface devices (e.g., 802.11 cards, 3G wireless modems, or 4G wireless modems). In some implementations, the input / output device may include a driver device configured to receive input data and send output data to other input / output devices, such as a keyboard, printer, and display device 860. In some examples, mobile computing devices, mobile communication devices, and other devices may be used.

[0118] In some implementations, at least part of the above approach can be realized by instructions that, when executed, cause one or more processing units to perform the above processing and functions. Such instructions may include, for example, script instructions, or interpreted instructions such as executable code, or other instructions stored in non-temporary computer-readable media. The storage device 830 may be implemented, for example, as a server farm or a collection of widely distributed servers on a network, or it may be implemented on a single computing device.

[0119] Figure 8 shows an example of a processing system, but the embodiments, functional operations, and processes of the subject matter described herein can be implemented in other types of digital electronic circuits, in tangible computer software or firmware, in computer hardware including the structures disclosed herein and their structural equivalents, or in one or more combinations thereof. The embodiments of the subject matter described herein can be implemented as modules of one or more computer programs, i.e., one or more computer program instructions coded on a tangible non-volatile program carrier for execution by or control of a data processing device. Alternatively, the program instructions can be coded into artificially generated propagating signals, such as machine-generated electrical, optical, or electromagnetic signals generated to code information to be transmitted to a receiving device suitable for execution by a data processing device. The computer storage medium may be a computer-readable storage device, a computer-readable storage board, or a random or serial access memory device, or one or more combinations thereof.

[0120] The term "system" may encompass all kinds of devices, equipment, and machines for data processing, including, for example, a programmable processor, a computer, or multiple processors or computers. A processing system may include, for example, a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a programmable general-purpose microprocessor or microcontroller, or other special-purpose logic circuit. In addition to hardware, a processing system may include code that creates the execution environment for the computer program of interest, such as processor firmware, a protocol stack, a database management system, an operating system, or code that constitutes one or more of these.

[0121] A computer program (which may also be called a program, software, software application, module, software module, script, or code) can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and can be deployed in any form, including standalone programs, modules, components, subroutines, or other units suitable for use in a computing environment. A computer program may, but may not, correspond to a file in a file system. A program can be stored in a single file dedicated to the program of interest, or in a set of interconnected files (e.g., a file storing one or more modules, subprograms, or parts of code), as part of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document). A computer program can be deployed to run on one computer, or on multiple computers located in one location or distributed across multiple locations and interconnected by a communication network.

[0122] The processes and logic flows described herein can be implemented by one or more programmable computers executing one or more computer programs, performing operations on input data, and generating outputs to perform their functions. These processes and logic flows can also be implemented by special-purpose logic circuits such as FPGAs, ASICs, or programmable general-purpose microprocessors or microcontrollers, and the devices themselves can also be implemented as such special-purpose logic circuits.

[0123] Computers suitable for executing computer programs may include, for example, general-purpose microprocessors, special-purpose microprocessors, or both, or any other type of central processing unit. Generally, a central processing unit receives instructions and data from read-only memory, random-access memory, or both. A computer generally comprises a central processing unit for issuing or executing instructions, and one or more storage devices for storing instructions and data. Generally, a computer has one or more mass storage devices for storing data, such as magnetic disks, magneto-optical disks, or optical disks, or is operably connected to mass storage devices for receiving or transmitting data, or both. However, a computer does not necessarily have to have such devices. Furthermore, a computer can be incorporated into another device, for example, a mobile phone, a personal digital assistant (PDA), a portable audio / video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device (e.g., a Universal Serial Bus (USB) flash drive).

[0124] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, examples of which include semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory can be complemented by or incorporated into special-purpose logic circuits.

[0125] To provide user interaction, embodiments of the subject matter described herein can be implemented on a computer equipped with a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal screen) monitor), and a keyboard and pointing device for the user to input into the computer (e.g., a mouse or trackball). Other types of devices can also be used to provide user interaction. For example, the feedback provided to the user can be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback, and the input from the user can be acoustic, spoken language, or tactile input. Furthermore, the computer can interact with the user by sending and receiving documents to and from the device the user is using (e.g., by sending a web page to a web browser on the user's device in response to a request received from a web browser).

[0126] Embodiments of the Subject Matter described herein may be implemented in a computing system that includes a backend component (e.g., a data server), a middleware component (e.g., an application server), a frontend component (e.g., a client computer with a graphical user interface or web browser that allows a user to interact with an implementation of the Subject Matter described herein), or any combination of one or more such backend, middleware, or frontend components. The components of this system may be interconnected by digital data communications of any form or medium, such as communication networks. Examples of communication networks include local area networks ("LANs") and wide area networks ("WANs"), such as the Internet.

[0127] A computing system can include clients and servers. Clients and servers are generally far apart from each other and typically communicate via a communication network. The client-server relationship arises from computer programs running on each computer and from the client-server relationship they have with each other.

[0128] While this specification contains many specific implementation details, these should not be interpreted as limitations on the scope of what can be claimed, but rather as descriptions of features that may be specific to particular embodiments. Certain features described herein as separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may be implemented individually or in any preferred combination in multiple embodiments. Furthermore, even if features are described above as acting in a particular combination and are initially claimed as such, one or more features may be excluded from the claimed combination, and the claimed combination may cover a partial combination or a variation of a partial combination.

[0129] Similarly, while calculations are shown in a specific order in the diagrams, this should not be understood as meaning that such calculations must be performed in the specific order or sequence shown, or that all illustrated calculations must be performed, in order to obtain the desired result. In certain situations, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system components in the above embodiments should not be understood as requiring such separation in all embodiments, and it should be understood that the above-mentioned program components and systems can generally be integrated into a single software product or packaged into multiple software products.

[0130] Furthermore, programs implementing various aspects of several embodiments may be accessed from a remote location (e.g., a server) via a network. Such data and / or programs may be transmitted via various machine-readable media, including but not limited to magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs and holographic devices; magneto-optical media; and hardware devices specifically configured to store or store and execute program code, such as application-specific integrated circuits (ASICs), programmable logic units (PLDs), flash memory devices, and ROM and RAM devices. Some embodiments may code instructions for causing one or more processors or processing units to perform specific steps in one or more non-temporary computer-readable media. Note that one or more non-temporary computer-readable media shall include volatile and non-volatile memory. Also note that alternative implementations are possible, including hardware implementations or software / hardware implementations. Hardware implementation functions may be implemented using ASICs, programmable arrays, digital signal processing circuits, etc. Therefore, the term “means” in any claim is intended to encompass both software and hardware implementations. Similarly, the term “computer-readable medium” as used herein includes software and / or hardware, or a combination thereof, into which a program of instructions is incorporated. With alternatives to these implementations in mind, it should be understood that the drawings and accompanying descriptions provide functional information necessary for a person skilled in the art to create program code (i.e., software) and / or manufacture circuits (i.e., hardware) for performing the required processing.

[0131] It should be noted that some embodiments may further relate to computer products comprising a non-temporary, tangible, computer-readable medium on which computer code for performing various computer-implemented operations is recorded. The medium and computer code may be specifically designed and constructed for the purposes of the techniques described herein, or may be of a type known to or available to those skilled in the art in the relevant field. Examples of tangible computer-readable media include, but are not limited to, magnetic media such as hard disks, floppy disks, and magnetic tapes; optical media such as CD-ROMs and holographic devices; magneto-optical media; and hardware devices specifically configured to store or store and execute program code, such as application-specific integrated circuits (ASICs), programmable logic units (PLDs), flash memory devices, and ROM and RAM devices. Examples of computer code include, for example, machine code generated by a compiler, and files containing high-level code executed by a computer using an interpreter. Some embodiments may be implemented in whole or in part as machine-executable instructions that may be within a program module executed by a processing unit. Examples of program modules include libraries, programs, routines, objects, components, and data structures. In a distributed computing environment, program modules may be physically located in a local, remote, or both environment.

[0132] Those skilled in the art will recognize that computing systems or programming languages ​​are not essential for carrying out the techniques described herein. Those skilled in the art will also recognize that some of the above elements may be physically and / or functionally separated into submodules or combined together.

[0133] In several embodiments, aspects of the techniques described herein (e.g., training a risk assessment model, using the risk assessment model to assess the risks associated with a commitment request, performing one or more (e.g., all) of the steps of the methods described herein) may be implemented using machine learning and / or artificial intelligence techniques.

[0134] Machine learning generally refers to the application of specific techniques (e.g., pattern recognition techniques and / or statistical inference techniques) by a computer system to perform a particular task. Machine learning techniques may be used to build a model based on sample data (e.g., "training data") and to validate the model using validation data (e.g., "test data"). The sample and validation data may be organized as a set of records (e.g., "observations" or "data samples"), each record representing a value for a specified data field (e.g., "independent variable," "input," "feature," or "prediction") and corresponding values ​​for other data fields (e.g., "dependent variable," "output," or "target"). Machine learning techniques may be used to train a model to infer output values ​​based on input values. If other data similar to or related to the sample data (e.g., "inference data") is presented, such a model may accurately infer unknown values ​​for the target in the inference dataset.

[0135] A feature of a data sample can be a measurable characteristic of the entity (e.g., a person, object, event, activity, etc.) represented by or associated with the data sample. The value of a feature can be a measurement of the corresponding characteristic of the entity or an instance of information about the entity. A feature can also have a data type. For example, a feature can have an image data type, a numerical data type, a text data type (e.g., a structured text data type or an unstructured ("free") text data type), a categorical data type, or any other suitable data type. Generally, if the set of values ​​that can be assigned to a feature is finite, then the data type of that feature can be said to be categorical.

[0136] As used herein, “model” may refer to any suitable model artifact produced by the process of using machine learning algorithms to fit a model to a particular training dataset. The terms “model,” “data analysis model,” “machine learning model,” and “machine learning model” are used herein to mean the same thing.

[0137] As used herein, “development” of a machine learning model may mean constructing a machine learning model. A machine learning model can be constructed by a computer using a training dataset. Therefore, “development” of a machine learning model may include training the machine learning model using a training dataset. In some cases (commonly referred to as “supervised learning”), the training dataset used to train a machine learning model may contain known outcomes (e.g., labels or target values) of individual data samples within the training dataset. For example, when training a supervised computer vision model to detect images of cats, the target value of a data sample in the training dataset may indicate whether or not the data sample contains an image of a cat. In other cases (commonly referred to as “unsupervised learning”), the training dataset does not contain known outcomes of individual data samples within the training dataset.

[0138] After development, the machine learning model may be used to generate inferences on the “inference” dataset. For example, after development, the computer vision model may be configured to distinguish between data samples containing images of cats and data samples that do not contain images of cats. As used herein, “deployment” of a machine learning model may mean using the developed machine learning model to generate inferences on data other than the training data.

[0139] "Artificial intelligence" (AI) generally encompasses any technique that demonstrates intelligence. Applications that demonstrate intelligence (e.g., machine execution software) may be referred to herein as "artificial intelligence applications," "AI applications," or "intelligent agents." Intelligent agents may demonstrate intelligence, for example, by perceiving, learning, and / or solving problems (e.g., by taking actions or making decisions that increase the likelihood of achieving a defined goal). Often, intelligent agents are developed by organizations, deployed on networked computer systems, and accessible to users within the organization. Intelligent agents are used to guide decision-making and / or control systems in a wide range of fields and industries, such as security, transportation, risk assessment and management, supply chain logistics, and energy management. Intelligent agents may include or use models.

[0140] Some non-exclusive examples of AI application types include inference applications, comparison applications, and optimization applications. An inference application may include any intelligent agent that generates inferences (e.g., predictions, forecasts, etc.) about the values ​​of one or more output variables based on the values ​​of one or more input variables. In some examples, an inference application may provide recommendations based on the generated inferences. For example, an inference application for a lending institution may infer the likelihood that a loan applicant will default on repayments of a requested loan and, based on that inference, recommend whether to approve the loan. A comparison application may include any intelligent agent that compares two or more possible scenarios. Each scenario may correspond to a set of latent values ​​of one or more input variables over a certain period of time. For each scenario, the intelligent agent may generate one or more inferences (e.g., about the values ​​of one or more output variables) and / or recommendations. For example, a comparison application for a lending institution may display predicted returns for a certain period of time only if the lending institution approves the loan application, and only if the predicted default risk is less than 20% (scenario #1), less than 10% (scenario #2), or less than 5% (scenario #3). An optimization application may include any intelligent agent that estimates the optimal values ​​for one or more variables of interest based on the values ​​of one or more input variables. For example, an optimization application for a lending institution may indicate the maximum loan amount that the institution will approve for a particular customer.

[0141] As used herein, “data analysis” may mean the process of analyzing data (e.g., using machine learning models, artificial intelligence, models, or techniques) to discover information, draw conclusions, and / or support decision-making. Types of data analysis include descriptive analysis (e.g., the process of describing information, trends, anomalies, etc., within a dataset), diagnostic analysis (e.g., the process of inferring why certain trends, patterns, anomalies, etc., exist in a dataset), predictive analysis (e.g., the process of predicting future events or outcomes), and prescriptive analysis (the process of determining or proposing courses of action).

[0142] Data analysis tools are used to guide decision-making and / or control systems in a wide range of fields and industries, such as security, transportation, risk assessment and management, supply chain logistics, and energy management. The processes used to develop data analysis tools suitable for performing specific data analysis tasks generally include the steps of data collection, data preparation, feature engineering, model generation, and / or model deployment.

[0143] Several embodiments Some embodiments may include any of the following:

[0144] (A1) A method for measuring paraspinal muscle group parameters, the method comprising: acquiring a plurality of two-dimensional magnetic resonance imaging (MRI) image slices 115 of a spine of interest; identifying one or more representative slices 215 from the plurality of MRI slices 115 by a slice selection model 210, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the spine of interest; segmenting each of the one or more representative slices by a muscle segmentation model 220, wherein the segmentation of each representative slice includes generating an image segmentation mask that includes a plurality of pixels in each representative slice corresponding to the paraspinal muscle group and excludes a plurality of pixels in each representative slice corresponding to metallic artifacts and / or other biological material of interest other than the paraspinal muscle group; and extracting one or more measurements of each of the one or more parameters of the paraspinal muscle group of interest by a measurement extraction model 230, based on the plurality of pixels included in the generated image segmentation mask of each of the one or more representative slices.

[0145] (A2) The method according to A1, wherein the slice selection model 210 is configured to identify each of the one or more representative slices 215 using a many-to-one slice-to-probability approach.

[0146] (A3) The method according to A1 or A2, wherein the muscle segmentation model 220 is trained to exclude the plurality of pixels corresponding to metal artifacts using a training dataset that includes one or more slices to which metal artifacts have been synthetically added.

[0147] (A4) The method according to any one of A1 to A3, wherein the one or more representative slices include a first representative slice located at or adjacent to the upper vertebral endplate in the biomechanical environment corresponding to the C7 vertebral body or the upper vertebral endplate in the biomechanical environment corresponding to the midpoint of the C3 vertebral body at the cervical level of the spine in question.

[0148] (A5) The method according to any one of A1 to A4, wherein the one or more representative slices include a second representative slice located in or adjacent to the apical segment at the thoracic level of the vertebra of the subject spine.

[0149] (A6) The method according to A5, wherein the apical segment of the vertebra is located in or adjacent to the T8, T7, T6, or T5 vertebral body.

[0150] (A7) The method according to any one of A1 to A6, wherein the one or more representative slices include a third representative slice located at or adjacent to the upper vertebral endplate in the biomechanical environment corresponding to the L4 vertebral body at the lumbar spine level of the subject.

[0151] (A8) The method according to any one of A1 to A7, wherein one or more of the paraspinal muscles of the subject include functional cross-sectional area (FCSA), fat infiltration (FI), fat distribution, Sobel index, and / or quality of the paraspinal muscles (PMQ).

[0152] (A9) The method of A8, further comprising characterizing the degree of sarcopenia in the subject based on the extracted measurements of one or more parameters.

[0153] (A10) The method according to any one of A1 to A7, wherein one or more parameters of the paraspinal muscle group of the subject include functional cross-sectional area (FCSA) and fat infiltration (FI), and the method further comprises predicting low back pain in the subject based on the extracted measurements of FCSA and FI.

[0154] (A11) The method according to any one of A1 to A7, wherein one or more parameters of the paraspinal muscles of the subject include fat distribution and fat infiltration (FI) within the posterior paraspinal muscles of the subject, and the method further comprises evaluating the status of at least one degenerative disease in the subject based on the extracted measurements of fat distribution and FI within the posterior paraspinal muscles of the subject.

[0155] (A12) The method according to A11, wherein the at least one degenerative disease includes spinal stenosis, facet joint degeneration, spinal instability, or spinal deformity.

[0156] (A13) The method according to any one of A1 to A7, wherein one or more parameters of the paraspinal muscle group of the subject include a first fat infiltration (FI) in the lumbar multifidus (MF) muscle group of the subject and a second FI in the erector spinae (ES) muscle group of the subject, and the method further comprises predicting the outcome of lumbar fusion surgery for the subject based on the extracted measurements of the first FI and the second FI.

[0157] (A14) The method of A13, wherein predicting the outcome of the lumbar spinal fusion surgery includes predicting an adverse effect on the outcome of the lumbar spinal fusion surgery based on the extracted measurement of the first FI being at least 60% and the extracted measurement of the second FI being at least 40%.

[0158] (A15) The method according to any one of A1 to A14, wherein the slice selection model includes a first step and a second step.

[0159] (A16) The method according to A15, wherein the first step of the slice selection model includes a two-dimensional convolutional neural network (CNN), a transformer, an autoencoder, and / or an encoder-decoder architecture.

[0160] (A17) The method according to A15 or A16, wherein the second step of the slice selection model includes a recurrent neural network (RNN), a transformer, a fully connected neural network, or a deep neural network.

[0161] (A18) The method according to A17, wherein the RNN has an iterative network architecture.

[0162] (A19) The method according to any one of A15 to A18, wherein the plurality of MRI slices comprises one or more sets of MRI slices corresponding to each of one or more vertebral levels of the vertebra in question, and identifying the one or more representative slices in the plurality of MRI slices comprises, by a first step of the slice selection model, generating a coding for each of the image slices in the plurality of image slices, and, for each of the one or more sets of MRI slices, by a second step of the slice selection model, selecting the representative slice located in or adjacent to a designated physiological region corresponding to the specific biomechanical environment at each vertebral level of the vertebra in question, wherein the selection of each representative slice at each vertebral level is based on the coding of the image slice in the plurality of image slices.

[0163] (A20) The method according to A19, wherein the selection of each representative slice at each vertebral level is further based on the order of the image slices in the plurality of image slices.

[0164] (A21) The method according to A19 or A20, wherein the selection of each representative slice at each vertebral level is further based on one or more contextual features that represent one or more attributes of the subject.

[0165] (A22) The method according to A15, wherein the slice selection model is a Siamese neural network model, the first stage of the slice selection model includes a twin neural network, and the second stage of the slice selection model is configured to determine the distance and / or similarity between slices included in the plurality of MRI slices and a reference slice.

[0166] (A23) The method according to A15, wherein the slice selection model is a Vision transformer model, the first stage of the slice selection model is configured to perform a linear projection of an input image slice, and the second stage of the slice selection model includes a plurality of classification heads.

[0167] (A24) The method according to any of A1 to A14, wherein the slice selection model includes one or more steps.

[0168] (A25) The method according to any of A1 to A24, wherein the muscle segmentation model includes a U-net-based convolutional neural network (CNN) model.

[0169] (A26) The method according to A25, wherein the U-net based CNN model comprises a plurality of first stages, each of which comprises a plurality of convolutional layers 452 and downsampling layers 454.

[0170] (A27) The method according to A23 or A26, wherein the U-net based CNN model includes a plurality of second stages, each of which includes a plurality of convolutional layers 462 and upsampling layers 464.

[0171] (A28) The method according to any one of A1 to A27, wherein at least a subset of the two-dimensional image slices of the target spine contains metallic artifacts.

[0172] (A29) The method of A28, wherein, for each representative slice, each image segmentation mask separates the plurality of pixels corresponding to the paraspinal muscles from the plurality of pixels corresponding to the metallic artifact and the plurality of pixels corresponding to biological material other than the paraspinal muscles.

[0173] (A30) The method according to A1, wherein the paraspinal muscle group contains fat, and the plurality of pixels in each representative slice corresponding to the paraspinal muscle group include one or more pixels corresponding to the contained fat.

[0174] (A31) The method of A2, wherein using a many-to-one slice-to-probability approach includes selecting one or more representative slices using each of the multiple image slices 115.

[0175] (A32) The method according to A1, wherein the muscle segmentation model 220 is trained to exclude the plurality of pixels corresponding to the metal artifact using transfer learning on a training dataset which includes one or more slices in which the true metal artifact is present.

[0176] (A33) A system for measuring paraspinal muscle group parameters, comprising: one or more processing units; acquiring multiple magnetic resonance imaging (MRI) two-dimensional image slices 115 of a target spine when performed by the one or more processing units; identifying one or more representative slices 215 of the multiple MRI slices 115 by a slice selection model 210, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine; and segmenting each of the one or more representative slices by a muscle segmentation model 220. The system comprises generating an image segmentation mask in which segmentation of each representative slice includes a plurality of pixels in each representative slice corresponding to the paraspinal muscles and excludes a plurality of pixels in each representative slice corresponding to metallic artifacts and / or the biological material of the subject other than the paraspinal muscles, and one or more storage devices that store instructions causing the system to perform calculations including extracting one or more measurements of one or more parameters of the paraspinal muscles of the subject based on the plurality of pixels in the generated image segmentation mask of each of the one or more representative slices using a measurement extraction model 230.

[0177] (A34) The system according to A33, further comprising a magnetic resonance imaging (MRI) scanner configured to generate the two-dimensional image slices of the spine of the subject.

[0178] (A35) When performed by one or more processing units of the system, the acquisition of multiple magnetic resonance imaging (MRI) two-dimensional image slices 115 of the target spine, the identification of one or more representative slices 215 of the multiple MRI slices 115 by a slice selection model 210, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine, and the segmentation of each of the one or more representative slices by a muscle segmentation model 220, wherein the segmentation of each representative slice is A computer-readable storage medium storing instructions configured to cause the system to perform calculations including generating an image segmentation mask that includes a plurality of pixels in each representative slice corresponding to the paraspinal muscles and excludes a plurality of pixels in each representative slice corresponding to metallic artifacts and / or the biological material of the subject other than the paraspinal muscles, and extracting one or more measurements of one or more parameters of the paraspinal muscles of the subject based on the plurality of pixels contained in the generated image segmentation mask of each of the one or more representative slices using a measurement extraction model 230.

[0179] term The expressions and terms used herein are for illustrative purposes only and should not be considered limiting.

[0180] Measurements, sizes, quantities, etc., may be presented in range form in this specification. Range form is provided for convenience and conciseness and should not be construed as a firm limitation on the scope of the invention. Therefore, range descriptions should be considered to specifically disclose all possible subranges and individual numerical values ​​within those ranges. For example, a range description such as 1 to 20 meters should be considered to specifically disclose subranges such as 1 meter, 2 meters, 1 to 2 meters, less than 2 meters, 10 to 11 meters, 10 to 12 meters, 10 to 13 meters, 10 to 14 meters, 11 to 12 meters, and 11 to 13 meters.

[0181] Furthermore, the connections between components or systems in the diagram are not limited to direct connections. Rather, data or signals between these components may be modified, reformatted, or otherwise altered by intermediate components. Additional or fewer connections may also be used. The terms “coupling,” “connection,” or “communication coupling” are understood to include direct connections, indirect connections through one or more intermediate devices, wireless connections, and so on.

[0182] References to “one embodiment,” “preferred embodiment,” “an embodiment,” “several embodiments,” or “multiple embodiments” in this specification mean that certain features, structures, characteristics, or functions described in relation to that embodiment are included in at least one embodiment of the present invention, and may be included in multiple embodiments. Furthermore, the occurrence of the above phrases in various parts of this specification does not necessarily refer to the same embodiment or multiple embodiments.

[0183] The use of certain terms in various parts of this specification is illustrative only and should not be construed as limiting. Services, features, or resources are not limited to a single service, feature, or resource; the use of these terms may refer to a group of related services, features, or resources that may be distributed or aggregated.

[0184] Furthermore, a person skilled in the art will recognize that (1) certain steps may be performed at their discretion, (2) the steps are not limited to the specific order described herein, (3) certain steps may be performed in a different order, and (4) certain steps may be performed simultaneously or in parallel.

[0185] As used herein and in the claims, the terms “approximately,” the phrase “approximately equal,” and other similar phrases (e.g., “X has a value of approximately Y” or “X is approximately equal to Y”) should be understood to mean that one value (X) is within a given range of another value (Y). Unless otherwise indicated, the given range may be 20%, 10%, 5%, 3%, 1%, 0.1%, or less than 0.1%.

[0186] Where used herein and in the claims, the indefinite articles “a” and “an” should be understood to mean “at least one” unless otherwise indicated. Where used herein and in the claims, the phrase “and / or” should be understood to mean “either one or both” of the combined elements, that is, elements that exist together in some cases and separately in others. Similarly, any multiple elements listed in “and / or” should be interpreted as “one or more” of those elements combined. Other elements may exist, at their discretion, other than those specifically identified by the “and / or” clause, whether or not they relate to the specifically identified elements. Thus, as a non-restrictive example, a reference to “A and / or B,” when used in conjunction with an open-ended expression such as “comprising,” may refer to A only in one embodiment (including, at their discretion, elements other than B), to B only in another embodiment (including, at their discretion, elements other than A), and to yet another embodiment to both A and B (including, at their discretion, other elements).

[0187] Where used herein and in the claims, “or” should be understood to have the same meaning as “and / or” as defined above. For example, “or” or “and / or” when separating items in a list should be interpreted as comprehensive, that is, including not only some or at least one of the elements in the list, but also multiple, and even optionally, additional items not in the list. Only terms that explicitly indicate the opposite, such as “only one of” or “exactly one,” or, where used in the claims, the term “consisting of,” should refer to including several or exactly one of the elements in the list. In general, the term “or” should be interpreted as an exclusive choice (i.e., “either one, but not both”) only when preceded by a phrase expressing exclusivity, such as “either one,” “one,” “only one,” or “exactly one.” Where used in the claims, “essentially consisting of” should have its usual meaning as it is in the field of patent law.

[0188] When used herein and in the claims, the phrase “at least one” when referring to a list of one or more elements should be understood to mean at least one element selected from any one or more elements in that list of elements, but not necessarily including at least one of each and all of the elements specifically listed in that list of elements, nor excluding any combination of elements in that list of elements. This definition also acknowledges that elements other than those specifically designated in the list of elements may exist at will, regardless of whether the phrase “at least one” relates to a specifically identified element. Therefore, as a non-restrictive example, “at least one of A and B” (or equivalently “at least one of A or B” or equivalently “at least one of A and / or B”) may, in one embodiment, refer to at least one (optionally including multiple) A in which B is absent (and optionally including elements other than B); in another embodiment, refer to at least one (optionally including multiple) B in which A is absent (and optionally including elements other than A); and in yet another embodiment, refer to at least one (optionally including multiple) A and at least one (optionally including multiple) B (and optionally including other elements).

[0189] The use of "including," "comprising," "having," "containing," "involving," and variations thereof means to include the items and additional items listed thereafter.

[0190] The use of ordinal terms such as "first," "second," and "third" to modify claim elements in the claims does not, in itself, imply that one claim element takes precedence over another, has priority, or is in order, nor does it imply a temporal order in which the methods are performed. Ordinal terms are used solely as labels to distinguish claim elements, to differentiate a claim element having a particular name from another element having the same name (except through the use of ordinal terms).

[0191] Specific embodiments of this subject matter have been described. Other embodiments are within the scope of the following claims. For example, the desired results can still be obtained even if the steps described in the claims are performed in a different order. As an example, the process shown in the accompanying drawings does not necessarily have to be performed in the specific order shown, or in any order, to obtain the desired results. In certain implementations, multitasking and parallel processing may be advantageous. Other steps or stages may be added, or steps or stages may be omitted from the described process. Therefore, other implementations are within the scope of the following claims.

[0192] Those skilled in the art will understand that the examples and embodiments described above are illustrative and not limit the scope of this disclosure. Any permutations, modifications, equivalents, combinations, and improvements that will become apparent to those skilled in the art who have read this specification and examined the drawings are intended to be included in the true spirit and scope of this disclosure. It should also be noted that any element of any claim may be arranged in different ways, including having multiple dependencies, configurations, and combinations.

[0193] While several aspects of at least one embodiment of the present invention have been described, those skilled in the art will understand that various changes, modifications, and improvements are readily conceivable. Such changes, modifications, and improvements are intended to constitute part of this disclosure and to be within the spirit and scope of the invention. Accordingly, the foregoing description and drawings are merely examples.

Claims

1. A method for measuring paraspinal muscle group parameters, wherein the method is To acquire multiple two-dimensional magnetic resonance (MRI) image slices of the target spine, The slice selection model identifies one or more representative slices from the plurality of MRI slices, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine. The muscle segmentation model is used to segment each of the one or more representative slices, wherein the segmentation of each representative slice is performed as follows: An image segmentation mask that includes a plurality of pixels corresponding to the paraspinal muscles in each of the aforementioned representative slices, and excludes a plurality of pixels corresponding to metallic artifacts and / or the subject biological material other than the paraspinal muscles in each of the aforementioned representative slices. Including generating, and The measurement extraction model extracts one or more measurements of one or more parameters of the target paravertebral muscle group based on the plurality of pixels contained in the generated image segmentation mask of each of the one or more representative slices. The method, including the method described above.

2. The method according to claim 1, wherein the slice selection model is configured to identify each of the one or more representative slices using a many-to-one slice-to-probability approach.

3. The method according to claim 1, wherein the muscle segmentation model is trained to exclude the plurality of pixels corresponding to the metal artifacts using a training dataset that includes one or more slices to which metal artifacts have been synthetically added.

4. The one or more representative slices mentioned above A first representative slice located at or adjacent to the biomechanically efficient upper vertebral endplate corresponding to the C7 vertebral body or the biomechanically efficient upper vertebral endplate corresponding to the center of the C3 vertebral body at the cervical level of the spine of the subject. The method according to claim 1, including the method described in claim 1.

5. The method according to claim 1, wherein the one or more representative slices include a second representative slice located in or adjacent to the apical segment at the thoracic level of the vertebra of the spine in question.

6. The method according to claim 5, wherein the apical segment of the vertebra is located in or adjacent to the T8, T7, T6, or T5 vertebral body.

7. The method according to claim 1, wherein the one or more representative slices include a third representative slice located at or adjacent to the upper vertebral endplate in the biomechanical environment corresponding to the L4 vertebral body at the lumbar spine level of the subject.

8. The method according to claim 1, wherein one or more parameters of the paraspinal muscle group of the subject include functional cross-sectional area (FCSA), fat infiltration (FI), fat distribution, Sobel index, and / or quality of the paraspinal muscle group (PMQ).

9. The method according to claim 8, further comprising characterizing the degree of sarcopenia in the subject based on the extracted measurements of one or more of the parameters.

10. The method according to claim 1, wherein one or more parameters of the paraspinal muscle group of the subject include functional cross-sectional area (FCSA) and fat infiltration (FI), and the method further comprises predicting low back pain in the subject based on the extracted measurements of FCSA and FI.

11. The method according to claim 1, wherein one or more parameters of the paraspinal muscle group of the subject include fat distribution and fat infiltration (FI) within the posterior paraspinal muscle group of the subject, and the method further comprises evaluating the status of at least one degenerative disease in the subject based on the extracted measurements of fat distribution and FI within the posterior paraspinal muscle group of the subject.

12. The method according to claim 11, wherein the at least one degenerative disease includes spinal stenosis, facet joint degeneration, spinal instability, or spinal deformity.

13. The method according to claim 1, wherein one or more parameters of the paraspinal muscle group of the subject include a first fat infiltration (FI) in the lumbar multifidus (MF) muscle group of the subject and a second FI in the erector spinae (ES) muscle group of the subject, and the method further comprises predicting the outcome of lumbar fusion surgery for the subject based on the extracted measurements of the first FI and the second FI.

14. The method according to claim 13, wherein predicting the outcome of the lumbar spinal fusion surgery includes predicting an adverse effect on the outcome of the lumbar spinal fusion surgery based on the extracted measurement of the first FI being at least 60% and the extracted measurement of the second FI being at least 40%.

15. The method according to claim 1, wherein the slice selection model includes a first step and a second step.

16. The method according to claim 15, wherein the first step of the slice selection model includes a two-dimensional convolutional neural network (CNN), a transformer, an autoencoder, and / or an encoder-decoder architecture.

17. The method according to claim 15, wherein the second step of the slice selection model includes a recurrent neural network (RNN), a transformer, a fully connected neural network, or a deep neural network.

18. The plurality of MRI slices include one or more sets of MRI slices corresponding to one or more vertebral levels of the target spine, Identifying one or more representative slices among the plurality of MRI slices is The first step of the slice selection model involves generating an encoding for each of the image slices in the plurality of image slices, and For each of the one or more sets of MRI slices, the second step of the slice selection model selects a representative slice that is located in or adjacent to a designated physiological region corresponding to the specific biomechanical environment at each vertebral level of the target vertebra. Includes, The selection of each representative slice at each vertebral level is based on the coding of the image slices among the plurality of image slices, The method according to claim 15.

19. The method according to claim 18, wherein the selection of each representative slice at each vertebral level is further based on the order of the image slices in the plurality of image slices.

20. The method according to claim 15, wherein the slice selection model is a Siamese neural network model, the first stage of the slice selection model includes a twin neural network, and the second stage of the slice selection model is configured to determine the distance and / or similarity between slices included in the plurality of MRI slices and a reference slice.

21. The method according to claim 15, wherein the slice selection model is a Vision transformer model, the first stage of the slice selection model is configured to perform linear projection of an input image slice, and the second stage of the slice selection model includes a plurality of classification heads.

22. The method according to claim 1, wherein the slice selection model includes one or more steps.

23. The method according to claim 1, wherein the muscle segmentation model includes a U-net-based convolutional neural network (CNN) model.

24. The method according to claim 1, wherein at least a subset of the two-dimensional image slices of the spine of the subject includes metallic artifacts.

25. The method according to claim 24, wherein for each representative slice, each image segmentation mask separates the plurality of pixels corresponding to the paraspinal muscles from the plurality of pixels corresponding to the metallic artifact and the plurality of pixels corresponding to biological material other than the paraspinal muscles.

26. The method according to claim 1, wherein the paraspinal muscle group contains fat, and the plurality of pixels in each representative slice corresponding to the paraspinal muscle group include one or more pixels corresponding to the contained fat.

27. The method according to claim 2, wherein using a many-to-one slice-to-probability approach includes selecting one or more representative slices using each of the multiple image slices.

28. A system for measuring paraspinal muscle group parameters, wherein the system One or more processing units, When executed by one or more processing units, To acquire multiple two-dimensional magnetic resonance (MRI) image slices of the target spine, The slice selection model identifies one or more representative slices from the plurality of MRI slices, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine. The muscle segmentation model is used to segment each of the one or more representative slices, wherein the segmentation of each representative slice is performed as follows: An image segmentation mask that includes a plurality of pixels corresponding to the paraspinal muscles in each of the aforementioned representative slices, and excludes a plurality of pixels corresponding to metallic artifacts and / or the subject biological material other than the paraspinal muscles in each of the aforementioned representative slices. Including generating, and The measurement extraction model extracts one or more measurements of one or more parameters of the target paravertebral muscle group based on the plurality of pixels contained in the generated image segmentation mask of each of the one or more representative slices. One or more memory devices that store instructions for causing the system to perform calculations including the above The system comprising the above.

29. The system according to claim 28, further comprising a magnetic resonance imaging (MRI) scanner configured to generate the two-dimensional image slices of the spine of the subject.

30. When executed by one or more processing units of the system, To acquire multiple two-dimensional magnetic resonance (MRI) image slices of the target spine, The slice selection model identifies one or more representative slices from the plurality of MRI slices, wherein each of the one or more representative slices is located in or adjacent to a designated physiological region corresponding to a specific biomechanical environment at each vertebral level of the target spine. The muscle segmentation model is used to segment each of the one or more representative slices, wherein the segmentation of each representative slice is performed as follows: An image segmentation mask that includes a plurality of pixels corresponding to the paraspinal muscles in each of the aforementioned representative slices, and excludes a plurality of pixels corresponding to metallic artifacts and / or the subject biological material other than the paraspinal muscles in each of the aforementioned representative slices. Including generating, and The measurement extraction model extracts one or more measurements of one or more parameters of the target paravertebral muscle group based on the plurality of pixels contained in the generated image segmentation mask of each of the one or more representative slices. A computer-readable storage medium that stores instructions configured to cause the system to perform an operation including the above.

Images