Method of ascertaining a muscle fiber type composition
A non-invasive ultrasound-based method with a classifier allows for rapid and efficient classification of muscle fiber types, addressing the limitations of invasive biopsies and costly equipment in determining muscle fiber composition.
Patent Information
- Application Number
- PCT/EP2026/063262
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-07
- Filing Date
- 2026-05-07
- Publication Date
- 2026-11-12
AI Technical Summary
Current methods for determining muscle fiber type composition are invasive, costly, and impractical for widespread use, particularly in sports performance diagnostics, due to their reliance on invasive biopsies or expensive equipment.
A non-invasive method using ultrasound imaging and a classifier to analyze muscle fiber types based on muscle-type-specific characteristics, enabling rapid, automated classification with reduced equipment and energy consumption.
Enables efficient, automated, and cost-effective determination of muscle fiber type composition, suitable for mobile and real-time applications, reducing the need for invasive procedures and specialized facilities.
Smart Images

Figure EP2026063262_12112026_PF_FP_ABST
Abstract
Description
[0001] Methods for determining muscle fiber type composition
[0002] The present invention relates generally to the field of medical image processing and pattern recognition, in particular to a non-invasive method for determining the muscle fiber type composition of an individual.
[0003] Determining an individual's muscle fiber types or composition is crucial in various fields, particularly sports medicine, rehabilitation, and veterinary research. Current methods for identifying muscle fiber types are predominantly invasive, especially hollow-needle biopsy. This procedure allows for direct tissue sampling from the muscle, which is then analyzed histologically or biochemically to determine the muscle fiber composition. Despite its high accuracy, this method has significant drawbacks. It is not only painful and invasive but also carries potential risks such as infection or local tissue damage. Furthermore, it requires medical personnel and cannot be readily implemented on a large scale or for routine measurements.
[0004] An alternative method is magnetic resonance spectroscopy, which allows for a non-invasive analysis of muscle fiber type composition. However, this technique requires expensive and complex medical equipment available only in specialized facilities. Its application also requires a trained medical professional and, due to high costs and limited availability, is impractical for many potential applications. In particular, current methods are hardly feasible in areas such as sports performance diagnostics, where a rapid, mobile, and cost-effective classification of muscle fiber types would be necessary.
[0005] Based on this, the object of the present invention is to provide a non-invasive method for determining the muscle fiber type composition of an individual, which can be carried out using compact and portable ultrasound imaging devices and involves reduced equipment, energy, and computing costs. This object is achieved by the subject matter of the independent claims. Further developments are found in the dependent claims.
[0006] According to the invention, a method for determining the muscle fiber type composition of an individual is provided with a classifier, wherein the classifier assigns data of a human muscle acquired by means of ultrasound imaging to muscle fiber types of a plurality of different muscle fiber types based on muscle type-specific characteristics of the muscle data, comprising the following process steps:
[0007] Acquisition of data acquired via ultrasound imaging of a muscle of the individual,
[0008] Processing the muscle data so that a uniform feature standardization is present in the data,
[0009] Extracting muscle-type-specific characteristics from the processed data, taking into account the different muscle fiber types exhibiting different ultrasound reflectivities based on the muscle-type-specific characteristics, and inputting the muscle-type-specific characteristics into the classifier, whereby the classifier provides a composition of the proportions of different muscle fiber types in the form of the muscle fiber type composition upon input of the muscle-type-specific characteristics.
[0010] In this context, a muscle fiber type is understood to be a specific muscle structure characterized by functional and structural features and which can be subdivided into different types. Possessing a muscle fiber type means that a muscle or muscle region of an individual has a specific muscle fiber type. The muscle fiber type composition is accordingly made up of varying proportions of the respective muscle fiber types.
[0011] According to the invention, the classifier assigns data acquired from a human muscle using ultrasound imaging to a plurality of different muscle fiber types based on muscle-type-specific characteristics. A classifier can be understood as a computer-implemented or computer-aided component that assigns an input, the muscle-type-specific characteristics, to one of several possible muscle fiber types based on specific criteria. Ultrasound imaging is a method that generates muscle-specific data from reflected ultrasound waves and enables a non-invasive visualization of internal tissue structures. Data acquired from a muscle using ultrasound imaging, as defined in the invention, can therefore be understood as digital measured values that can be obtained with an ultrasound imaging device.A typical ultrasound imaging device comprises a transmitter configured to send an ultrasound signal into an object and a receiver configured to detect the signal reflected from the object. In this case, the data obtained through ultrasound imaging can be image data in the form of ultrasound images and / or radio frequency data. The data thus provides information about the tissue structure of the muscle and its composition. Muscle type-specific features are characteristic properties relevant for distinguishing different muscle fiber types and can be extracted from the acquired data. This allows for a non-invasive method for identifying muscle fiber types and subsequently determining their composition.
[0012] The procedure is designed to involve acquiring data from an individual's muscle using ultrasound imaging. Acquisition refers to the process of recording and processing data from a measurement. A muscle is a biological tissue responsible for the movement of body structures and consists of a combination of different muscle fiber types. An individual is a living organism that possesses muscles whose fiber types are to be classified.
[0013] The process further includes data preparation, where data preparation refers to the systematic processing of the collected data to prepare it for subsequent classification. Feature standardization describes the transformation of the features into a uniform range of values, thereby compensating for differences in scaling, measurement, or dispersion. This inventive use of standardized data preparation ensures that all features have comparable numerical properties, thus reducing biases caused by different value ranges or statistical fluctuations. This improves the robustness and accuracy of the classifier, as it can make consistent decisions regardless of the original distribution of the data.
[0014] According to the invention, it is further provided that muscle-type-specific features are extracted from the processed data, taking into account the different muscle fiber types exhibiting varying ultrasound reflectivities. Extraction means the targeted selection and calculation of features that are significant for distinguishing between different muscle fiber types. This allows relevant information to be filtered from the processed data in order to perform the subsequent classification based on the muscle-type-specific features. Muscle-type-specific features are therefore characteristic properties that are relevant for distinguishing between different muscle fiber types.
[0015] Ultrasound reflectivity describes the strength of reflected ultrasound waves as a function of tissue structure and density. Different muscle fiber types exhibit varying reflection behavior towards the ultrasound waves used in ultrasound imaging. This differing reflection behavior is based on structural and compositional differences between muscle fiber types, which alter the tissue microstructure and thus the spatial distribution of acoustic impedances at the subcellular and cellular levels. Therefore, physiological and biological properties of the muscle fibers are captured by these different reflection behaviors. Such differences influence the number of effective scattering centers, the scattering cross-section, the backscattering properties of the ultrasound signal, its spectral signal components, and the attenuation of ultrasound waves in muscle tissue.
[0016] Even if individual muscle fibers lie below the lateral resolution of clinical ultrasound systems, these microscopic structural and compositional differences result in measurable changes in the collective scattering properties of the muscle tissue. This leads to systematic differences in echogenicity, texture characteristics, and radiofrequency-based spectral parameters, enabling the differentiation of muscles with different dominant fiber type compositions.
[0017] The influence of these different ultrasound reflectivities on the data is extracted in the form of muscle-type-specific characteristics. This extraction allows for the targeted consideration of the different reflection properties of the muscle fiber types and makes the fiber-type-specific information contained in the ultrasound data usable for subsequent classification.
[0018] The extracted muscle-type-specific characteristics are then entered into the classifier. The classifier, in response to this input, generates a composition of the proportions of different muscle fiber types, expressed as the muscle fiber type composition. Input into the classifier means that the determined muscle-type-specific characteristics are transferred as a structured dataset, for example, as a vector. Based on these characteristics, the classifier makes a classification decision. This allows for an objective and quantitatively evaluable statement about the affiliation of muscle fibers to a specific muscle fiber type, thus enabling efficient and non-invasive muscle analyses. The muscle fiber type composition in the muscle data is comprised of the proportions of the different muscle fiber types.It is used to indicate the composition of a muscle, for example based on the percentage of different muscle fiber types.
[0019] The method according to the invention advantageously enables a rapid, computationally efficient, and automated classification of muscle fiber type composition based on non-invasively acquired ultrasound data. By avoiding invasive procedures such as biopsies, not only is the burden on the individual being examined significantly reduced, but the need for medically trained personnel is also considerably minimized. Compared to complex imaging techniques such as magnetic resonance spectroscopy, the method according to the invention is characterized by significantly lower energy and resource consumption. In particular, by subdividing the method according to the invention and separating the classifier, the computational effort can be locally limited, and decentralized evaluation on compact hardware can be implemented.
[0020] By inputting the already extracted muscle-type-specific features into the classifier instead of the complete raw ultrasound data, the amount of data to be transferred between feature extraction and classification is significantly reduced. This results in more efficient use of available computing power, reduced storage requirements, and lower bandwidth consumption during data processing and transmission. The method is therefore also suitable for mobile, real-time applications with limited system resources, for example, in the field, in rehabilitation settings, or in competitive sports.
[0021] According to a further development of the invention, the majority of muscle fiber types are designed to include a fast-twitch and a slow-twitch muscle fiber type, which differ in their contraction speed. Thus, upon inputting the muscle-type-specific characteristics into the classifier, a muscle fiber composition containing a proportion of both fast-twitch and slow-twitch muscle fiber types is generated. Contraction speed describes the rate at which a muscle fiber can shorten or generate force. A fast-twitch muscle fiber type is a muscle type that exhibits high force generation with a short contraction time, while a slow-twitch muscle fiber type has a lower contraction speed but higher fatigue resistance.
[0022] Slow-twitch muscle fibers are also known as type I fibers, while fast-twitch muscle fibers are known as type II fibers. Slow-twitch muscle fibers are typically used for sustained strength activities such as holding and carrying loads and primarily meet their energy needs through oxidative metabolism. Fast-twitch muscle fibers, on the other hand, are designed for rapid, powerful movements and primarily meet their energy needs through glycolysis, which explains the higher glycogen content of these fibers.
[0023] Slow- and fast-twitch muscle fibers differ structurally and compositionally in a way that manifests itself in their ultrasound reflectivities. Slow-twitch muscle fibers typically exhibit a higher mitochondrial and capillary density, increased myoglobin content, and greater intramuscular lipid content, which is why they appear macroscopically as red muscle fibers. In contrast, fast-twitch muscle fibers usually have larger fiber cross-sections, a higher glycogen content, and a lower oxidative organelle density, and appear macroscopically as white muscle fibers.
[0024] Due to the higher organelle and lipid density of slow-twitch muscle fibers, these fibers have a greater number of effective scattering centers in the muscle tissue. With an increasing number of scattering centers, the backscatter coefficient and the mean image brightness also increase, so that muscle tissue dominated by slow-twitch muscle fibers tends to appear more echogenic (i.e., brighter) in ultrasound images and exhibits correspondingly higher values in echogenicity and texture parameters, as well as in radiofrequency-based spectral parameters.
[0025] In contrast, the higher glycogen content of fast-twitch muscle fibers leads to a correspondingly increased intramuscular water retention, as glycogen binds water in the muscle. Since the acoustic impedance difference between water and the surrounding tissue structures is relatively small, and water therefore reflects sound less strongly than denser tissue components, this results in a hypoechoic, i.e., darker, appearance of the muscle tissue in the ultrasound image. Muscle areas dominated by fast-twitch muscle fibers therefore tend to appear darker in the ultrasound image.
[0026] These two effects, the stronger backscattering in slow-contracting muscle fibers due to the higher organelle and lipid density on the one hand, and the hypoechoic appearance in fast-contracting muscle fibers due to the higher glycogen content on the other, work together in a complementary way with regard to the differentiation of muscle fiber types and enable a reliable separation of muscles with different dominant fiber type composition based on the ultrasound reflectivities.
[0027] According to a further development of the invention, the data are image data in the form of ultrasound images and / or radio frequency data. Radio frequency data are the raw signals or raw data acquired from an ultrasound probe, the receiver of an ultrasound imaging device, prior to image processing. Image data can be understood as digitized visual representations obtained through ultrasound imaging. Ultrasound images are two-dimensional representations of internal tissue structures generated by the reflection of sound waves. The radio frequency data can be converted into image data by means of processing or reconstruction. Data within the meaning of the invention can therefore be, on the one hand, the radio frequency data acquired directly by the receiver of the ultrasound imaging device and, on the other hand, the image data that can be generated by processing the radio frequency data.This use of the data type according to the invention enables a differentiated analysis of tissue structures by utilizing both the visual texture and the spectral properties of the ultrasound reflections. This allows for higher diagnostic accuracy, as relevant features can be extracted from both the image and the raw signal level.
[0028] According to a further embodiment of the invention, the data consists of image data obtained from radio frequency data using time gain compensation combined with a Hilbert transform. Time gain compensation can be understood as a signal processing technique that compensates for the attenuation of sound waves with increasing depth to ensure uniform image brightness across different tissue layers. A Hilbert transform is a mathematical method for converting a real signal into an analytical signal containing phase and amplitude information. This inventive use of signal processing improves image quality by optimizing the contrast characteristics and depth dependence of the signals. Simultaneously, both the radio frequency data and the image data are obtained from the common raw signal. The image data does not need to be stored permanently.It is sufficient to store the high-frequency data, since the image data can be obtained through the operations shown above.
[0029] One embodiment of the invention provides that the muscle is a calf muscle and / or a thigh muscle, the majority of muscle fiber types differ in their respective contraction speeds, wherein the majority of muscle fiber types include a fast-contracting muscle fiber type and a slow-contracting muscle fiber type, and upon input of the muscle-type-specific characteristics into the classifier, the muscle fiber type composition containing the proportions of fast-contracting and slow-contracting muscle fiber types is provided. The respective proportions of muscle fiber types can vary accordingly from 0 to 100%. This use according to the invention with regard to specific muscle groups ensures that the measurements are applied to muscle groups that are of particular relevance for athletic or rehabilitative issues.This allows the method to be used specifically for the analysis of muscle performance, training adaptations and individual muscular characteristics.
[0030] In a further development of the invention, it is intended that the step of acquiring the data of the individual's muscle will instead be:
[0031] Acquisition of data of a muscle of an individual acquired by means of ultrasound imaging, wherein the data were acquired by aligning a transducer of an ultrasound imaging device longitudinally to the muscle fibers of the muscle in a time-dependent manner during a contraction of the muscle, so that the data show a time-dependent course of the muscle fibers in the form of an image sequence, and the subsequent steps are instead:
[0032] Processing the image sequence so that a uniform feature standardization exists in the image sequence, and
[0033] Extracting muscle-type-specific features from the processed image sequence, where a muscle-type-specific feature is the muscle's contraction velocity. A longitudinal direction can be understood as the orientation along the main axis of the muscle fibers, thus enabling the most detailed possible representation of the muscle fiber structure. An image sequence is a temporally ordered sequence of individual images that together provide a dynamic representation of muscle activity. Contraction velocity is a measure of the change in muscle length over time and can be used to determine muscle fiber types. This use of the measurement setup according to the invention ensures that not only static structures are analyzed, but that dynamic contraction processes are directly captured from the ultrasound data.This allows training and rehabilitation measures to be specifically tailored to individual muscular characteristics by determining functional muscle characteristics with high precision.
[0034] According to a further development of the invention, the method comprises the following additional steps after the step of acquiring data of a muscle of an individual acquired by means of ultrasound imaging:
[0035] Combining muscle data with data obtained via electromyography of the muscle, and in the step of extracting muscle-type-specific characteristics from the processed data:
[0036] Extracting muscle-type-specific features from the processed data, where a muscle-type-specific feature is the contraction velocity of the muscle.
[0037] Electromyography can be understood as a measurement technique for recording the electrical activity of a muscle during contraction or relaxation. Combining data refers to the process of merging different data types into a unified analysis model. This inventive use of combining image-based and electrophysiological data enables a more precise determination of muscle fiber types by incorporating both structural and functional information into the classification. This allows for a more comprehensive analysis of muscular properties, improving the differentiation of muscle fiber types and forming the basis for more precise training or therapy planning.According to a further embodiment of the invention, the classifier is designed to determine the muscle fiber type composition with spatial resolution, and the step of acquiring the data of the individual's muscle is instead:.
[0038] Acquisition of data acquired by ultrasound imaging of a muscle of an individual, wherein the data were acquired by aligning a transducer of an ultrasound imaging device transversely to the muscle fibers of the muscle, so that the data perpendicular to a skin surface of the individual show the depth of the muscle, and the step of inputting the muscle type-specific features instead:
[0039] Input of muscle type-specific characteristics into the classifier, whereby the classifier, in response to the input of the muscle type-specific characteristics, provides the muscle fiber type composition in an area of the muscle.
[0040] A transverse direction can be understood as a measurement direction that runs perpendicular to the main axis of the muscle fibers, thus enabling detailed depth measurement. This use of the measurement arrangement according to the invention allows for a three-dimensional analysis of the muscle fiber type composition by performing not only a global classification of the muscle but also a differentiated evaluation of the muscle fiber type distribution. Areas can be analyzed from the data. The data are fed to the classifier, which then provides the spatially resolved muscle fiber type composition in a region of the muscle. The classifier knows which region of the muscle corresponds to a corresponding region in the muscle data. Accordingly, the classifier is able to provide the spatially resolved muscle fiber type composition based on this data.
[0041] The entire muscle itself is thus composed, for example, of a certain proportion of fast-twitch and slow-twitch muscle fiber types. The method described above allows for the simple tracking of training progress by monitoring the development of muscle composition over time. This enables targeted training to increase the proportion of fast-twitch muscle fibers, for instance. One embodiment of the invention provides for the extraction of muscle-type-specific characteristics using a grayscale matrix and / or statistical features and / or edge or flank detection and / or Fourier or wavelet transformation and / or principal component analysis.
[0042] A grayscale matrix can be understood as a matrix that stores intensity values of an image and is used for texture analysis. The grayscale matrix captures the echogenicity differences and texture features contained in the ultrasound data, which, as explained above, are due to the different scattering and backscattering properties of muscle tissue with varying fiber compositions. Texture parameters such as contrast, correlation, homogeneity, and entropy, which serve as muscle-type-specific characteristics, can be derived from the grayscale matrix.
[0043] Statistical characteristics include parameters such as mean, variance, skewness, and kurtosis of the pixel intensities of the ultrasound data or selected image areas. These parameters directly reflect the systematic intensity differences caused by the varying ultrasound reflectivities of different muscle fiber types and are therefore suitable as muscle-type-specific characteristics.
[0044] Edge detection is an image processing method used to identify transition zones between different structures. Edge detection is applied to image data, while flank detection is used for radio frequency data. These methods detect transitions between tissue regions with different echogenicity or different reflection amplitudes, which in turn result from fiber-type-specific differences in the number and distribution of effective scattering centers.
[0045] A Fourier or wavelet transform describes a mathematical method for decomposing a signal into its frequency components. These transformations make accessible the spectral signal components contained in the ultrasound data, the distribution of which, as explained above, is influenced by the scattering cross-section and the attenuation behavior of the muscle tissue, and thus by the fiber type composition. The resulting spectral parameters serve as muscle-type-specific characteristics.
[0046] Principal component analysis is a dimensionality reduction technique that identifies the most important axes of variation in a dataset. Applied to ultrasound data, it extracts the components that explain the largest proportion of the variance and thus represent the distinguishing characteristics between muscles with different fiber compositions.
[0047] This extraction of muscle-type-specific features according to the invention enables a robust and scalable analysis of ultrasound data and ensures that the signal characteristics arising from the different ultrasound reflectivities of different muscle fiber types are made available in a suitable form for subsequent classification.
[0048] As part of a further development of the invention, it is envisaged that the method in the step of processing the muscle data, so that a uniform feature standardization is present in the data, comprises the following process steps:
[0049] Segmenting the muscle data,
[0050] Removing a background, filtering the muscle data, and normalizing the muscle data.
[0051] Data segmentation refers to the separation of relevant image areas from irrelevant areas to enable targeted data analysis. Background removal and muscle data filtering describe the elimination of non-information-bearing image segments that are irrelevant for classification. This use of signal processing according to the invention reduces artifacts and irrelevant signal components, thereby improving signal quality. This enables a more stable and reliable classification of muscle fiber type composition later on. The method further includes normalizing the muscle data. Normalization describes a mathematical process in which values are converted into a standardized range to ensure comparability and robustness in subsequent processing steps.This allows fluctuations resulting from different recording conditions or individual anatomical features to be compensated for, so that the subsequent classification is not affected by irrelevant variations.
[0052] A further embodiment of the invention provides that segmentation is performed using feature-based image processing or a trained artificial neural network. Feature-based image processing can be understood as an algorithmic method for recognizing specific patterns or structures in image data based on mathematical models. A trained artificial neural network is a machine learning model that has been trained on training datasets to automatically recognize and classify relevant patterns. This use of the segmentation method according to the invention enables higher precision and automation in the identification of muscle fiber types by using either rule-based feature analyses or data-driven learning methods. This allows for increased processing speed and a reduction in dependence on manual segmentation steps.
[0053] According to a further development of the invention, the method in the step of processing the muscle data, so that a uniform feature standardization is present in the data, comprises the following additional process steps:
[0054] Trimming the data so that data sections not required for extracting muscle-type-specific characteristics are removed,
[0055] Adjusting the depth of focus in the data,
[0056] Adjusting the contrast and / or sharpness of the data, and
[0057] Inversion of the gray values of the data.
[0058] Data cropping can be understood as a process in which image areas irrelevant for subsequent feature extraction are removed to increase processing efficiency. Depth of focus describes the depth plane at which ultrasound images are optimally focused, ensuring that relevant structures are sharply rendered. This depth of focus can be adjusted in the data after acquisition. Adjusting the contrast and / or sharpness of the data refers to optimizing its visual appeal through methods such as histogram balancing or contrast-limited adaptive histogram balancing, thereby improving image quality. Inverting the grayscale values allows for improved image visualization.This use of image processing according to the invention increases signal quality and improves the recognition of muscle-type-specific features by removing interfering elements, enhancing contrasts, and optimizing the data for further processing. This increases the robustness of the classification method to varying recording conditions.
[0059] According to a further embodiment of the invention, the classifier provides a muscle fiber type space in which multiple classes for different muscle fiber type compositions are defined. Based on the muscle-type-specific features extracted from the data, the classifier assigns the data to one of the classes, and the muscle fiber type composition is determined based on the distances of the muscle data points to the centers of the classes. A muscle fiber type space can be understood as a multidimensional representation in which different muscle fiber type compositions are organized as separate classes. Each class corresponds to a muscle fiber type composition. The distance to a center of the class is a numerical metric for evaluating the similarity of a data point to a class.This use of a clustering approach according to the invention, for example by a k-means algorithm or a Gaussian mixture models algorithm, enables a robust and statistically sound classification of muscle fiber type composition. This quantifies the probability of class assignment based on similarity to representative cluster centers, the classes.
[0060] In a further development of the invention, it is intended that the method is additionally designed with an artificial neural network trained to determine the muscle fiber type composition, in the form of a network model. This network model includes a classifier, has been trained using a training dataset, and contains data from multiple individuals corresponding to a plurality of different muscle fiber type compositions. Upon input of the muscle type-specific features, the network model provides the muscle fiber type composition using the classifier. An artificial neural network can be understood as a machine learning model consisting of multiple layers of artificial neurons that can learn patterns in data through training. A training dataset is a set of annotated example data used to train the neural network.The majority of people represent a representative sample, which contributes to the generalization of the model to different individuals. This use of a trained neural network in the form of the network model, according to the invention, enables automatic and adaptive classification of muscle fiber type compositions by allowing the network model to learn from a large number of data points. This achieves high accuracy by optimizing the network architecture to recognize and reliably classify differences in muscle-type-specific characteristics. Likewise, the use of the trained artificial neural network significantly reduces the memory and energy requirements for applying the method.
[0061] Alternatively, in a further development of the invention, the trained artificial neural network can be provided with a network model that includes a k-nearest neighbor, a random forest, or a support vector machine algorithm, using the training dataset described above. A k-nearest neighbor algorithm can be understood as an instance-based learning method that assigns new data points to a specific muscle fiber composition based on their similarity to previously classified neighbors. A random forest algorithm is an ensemble method based on decision trees that enables robust classification of muscle fiber composition through majority decisions from several randomly generated trees. A support vector machine is a distance-maximizing classifier that determines an optimal interface between different muscle fiber compositions.
[0062] According to a further development of the invention, the training dataset is augmented by mirroring, rotating, translating, scaling, and / or splitting the data of the majority of people. Mirroring can be understood as a transformation in which the image is reflected around an axis. Rotation describes a rotation of the image data by a specific angle. Translation refers to a displacement of the image data in a defined direction. Scaling is a change in image size while maintaining the proportions, while splitting represents a segmentation of the data to generate additional training examples. This use of data augmentation according to the invention increases the variability of the training dataset, thereby enabling the neural network to achieve more robust generalization to new, unknown data.This improves the performance of the classifier by taking into account a greater variety of possible image variations.
[0063] According to a further embodiment of the invention, the network model comprises an extraction module and a pooling module. The extraction module extracts the muscle-type-specific features from the processed data. The extraction module is linked to the pooling module, which reduces the muscle-type-specific features. The pooling module is linked to the classifier, and the classifier analyzes the reduced muscle-type-specific features and provides the muscle fiber type composition based on this analysis. Thus, in addition to the classifier, the network model includes an extraction module and a pooling module and can therefore also be used to extract the muscle-type-specific features. An extraction module can be understood as a component that extracts relevant features from input data.A pooling module is a functional unit that reduces the number of muscle-type-specific features to increase data processing efficiency. The classifier analyzes the processed features and determines class membership based on learned patterns. In this sense, the network model extracts the muscle-type-specific features and provides the muscle fiber type composition. This use of a modular network architecture according to the invention improves the efficiency and accuracy of feature extraction and classification by enabling a stepwise reduction of the data volume while simultaneously preserving relevant information. This optimizes processing speed while maintaining classification accuracy.
[0064] According to one embodiment of the invention, the muscle-type-specific features are extracted in a penultimate layer. These features are then reduced to a single dimensionality by a pooling module forming the final layer. Based on these extracted features, the classifier assigns the data to one of the classes of different muscle fiber compositions. The probability of a given muscle fiber type being present in the data is determined by the distances of the muscle data to the centers of the respective classes. In this case, the classifier is a k-means algorithm. The network model comprises the artificial neural network in the form of a VGGlö model, the pooling module in the form of a principal component analysis, and the classifier.By representing the muscle-type-specific characteristics in a lower-dimensional space using Principal Component Analysis, the calculations are accelerated and improved interpretability of the data is achieved.
[0065] A further embodiment of the invention provides that, in the case of a network model with a k-nearest neighbor algorithm, the method additionally outputs a confidence value upon input of the muscle-type-specific characteristics into the classifier. According to a further embodiment of the invention, in a k-means algorithm, the absolute difference between a first vector of a first class defined by the muscle-type-specific characteristics and a second vector of a second class defined by the muscle-type-specific characteristics corresponds to a confidence value. Thus, the method according to the invention also provides that a probability is output indicating the presence of the respective muscle fiber type composition. This probability scales accordingly with the distance of the data from the respective class.
[0066] In this context, a further embodiment of the invention provides that the confidence value is above a predetermined value to indicate with sufficient probability that the muscle fiber type composition is present in the data. A confidence value can be understood as a numerical assessment that reflects the certainty of the classification decision. This use of a probability measure in classification, as described in the invention, enables a transparent evaluation of the reliability of the prediction. This allows the decision certainty to be quantified, so that alternative evaluation methods can be used or a remeasurement can be performed if the confidence is low.
[0067] In a further development of the invention, the network model for extracting muscle-type-specific features from the image data is provided to include edge or flank detection. Edge or flank detection can be understood as a method for identifying abrupt signal transitions in order to extract structural information. This use of a feature extraction method according to the invention improves the data analysis by recognizing important tissue boundaries and muscle fiber structures with high precision. This creates the basis for a more reliable differentiation of muscle fiber types.
[0068] A further embodiment of the invention provides that muscle-type-specific features are extracted from ultrasound image data using edge detection and / or from radio frequency data using edge detection. Edge detection describes the identification of abrupt intensity changes in an image, while edge detection enables the analysis of transition regions in a continuous signal. This use of specific extraction methods according to the invention ensures that both the visual image structure and the spectral signal dynamics are incorporated into the feature analysis. This increases the robustness of the classification by combining different feature representations.
[0069] In a further development of the invention, it is intended that the network model for extracting the muscle-type-specific features of the image sequence comprises a recurrent neural network or a long-short-term memory network. A recurrent neural network can be understood as a network architecture that processes temporal dependencies in sequential data. A long-short-term memory network is a specialized form of a recurrent network that has a memory structure for long-term information retention. This use of networks with temporal information processing, as described in the invention, enables a more precise analysis of changes in muscle contractions over time. This improves the detection of dynamic processes, which is crucial for differentiating muscle fiber types based on temporal patterns.
[0070] According to a further development of the invention, the largest value of the output of a final layer determines membership in a muscle fiber type composition. A final layer can be understood as the concluding processing layer of a neural network, which makes the final classification decisions. The largest value of the output denotes the highest numerical output of an activation vector representing the most probable muscle fiber type composition. This use of a maximization principle according to the invention ensures a unique class assignment by deriving the most probable class directly from the network output. This enables consistent and efficient decision-making without the need for additional calculations to weight alternatives.
[0071] According to a further embodiment of the invention, the classifier is provided with a softmax function that indicates the probability of a particular muscle fiber type composition being present in the muscle data. A softmax function can be understood as a mathematical transformation that converts raw values into normalized probabilities across multiple classes of muscle fiber type compositions. This use of a probability distribution according to the invention ensures a clear and comprehensible classification output by guaranteeing that the sum of the class assignments always equals one. This increases the comparability of classification decisions, allowing uncertainties to be quantified and classification results to be interpreted on a probabilistic basis.
[0072] As part of a further development of the invention, it is provided that the method includes the following additional process step: Representing the areas from which the muscle-type-specific features were extracted in the form of a heat map.
[0073] A heatmap can be understood as a visual representation that indicates the relevance of different image areas for a classification decision through color coding. This use of an image-based visualization method according to the invention improves the interpretability of the classification results by making it clear which image regions contribute significantly to the classification of muscle fiber type composition. This increases the transparency of the neural network, which is particularly important for medical applications where explainable models are preferred.
[0074] A further embodiment of the invention provides that the extraction module comprises a plurality of two- or three-dimensional convolution layers, the pooling module comprises a number of pooling layers corresponding to a number of convolution layers, and the convolution layers are alternately linked to the pooling layers. A two- or three-dimensional convolution layer can be understood as a neural network layer that extracts local features from image data by means of convolution operations. A pooling layer is a functional unit that reduces the number of features by selecting representative values from a group of neighboring pixels. This inventive use of a stepwise convolution and pooling structure ensures efficient feature processing by progressively generating higher levels of abstraction without losing the relevant features for muscle fiber type determination.This enables in-depth analysis of image data while minimizing the computational load.
[0075] The invention also relates to a data processing device for carrying out the process steps listed above. A data processing device can be understood as an electronic unit designed for acquiring, processing, feature extraction, and classifying muscle data. The device performs automated calculations and data manipulations to obtain structured information from input values.According to a preferred embodiment of the invention, the data processing device additionally comprises an electronic circuit, wherein the electronic circuit is configured to acquire data acquired by means of ultrasound imaging in the form of high-frequency data of a muscle of an individual and to process the high-frequency data of the muscle, a chip connected to the electronic circuit for signal transmission and comprising the classifier, wherein the chip is configured to extract the muscle-type-specific features from the processed high-frequency data, and the classifier provides the muscle fiber type composition upon input of the muscle-type-specific features.
[0076] According to the invention, the data processing device comprises an electronic circuit for executing the steps of acquiring data acquired by ultrasound imaging in the form of high-frequency data from an individual's muscle and processing this muscle data. An electronic circuit can be understood as an electronic circuit that performs computational operations and signal processing. The electronic circuit is therefore passive in the sense that no active software controls it. Its effect on the data is determined by the components integrated into the electronic circuit. In this context, a further development of the invention intends for the electronic circuit to be an integrated circuit.In this context, an integrated circuit is understood to be a circuit that corresponds to an electronic circuit realized in a compact semiconductor structure, which performs arithmetic operations and signal processing.
[0077] Furthermore, the data processing device includes a chip for performing the steps of extracting muscle-type-specific features from the processed data and delivering the muscle fiber type composition using the classifier. A classifier can be understood as a processor-based component that assigns an input to a specific muscle fiber type based on learned patterns. The classifier is thus contained within the chip as an embedded system. An embedded system is a specialized computing unit integrated into a device and optimized for a specific task, often operating with limited computing resources. Extracting muscle-type-specific features refers to the process of identifying and calculating relevant structural or texture-based properties that are important for distinguishing between different muscle fiber types.Following extraction, these features are used to determine and output the muscle fiber type composition. This use of the data processing device according to the invention enables efficient real-time analysis of muscle data with high classification accuracy by performing the signal processing directly in an optimized system, thereby minimizing computational effort and delays.
[0078] In a further development of the invention, it is intended that the chip be a microprocessor or a field-programmable gate array (FGA). A microprocessor can be understood as a unit for an embedded system that handles data processing and the execution of computational operations. A field-programmable gate array is a reconfigurable digital circuit that enables application-specific signal processing. This use of an embedded implementation according to the invention enables a fast and energy-efficient calculation of the probability with which the muscle fiber type composition is present in the data by utilizing optimized hardware structures for image analysis. This improves the real-time capability of determining the muscle fiber type composition, thereby making mobile and cost-effective applications of the invention feasible.
[0079] According to a further development of the invention, the signal processing in the electronic circuit is performed using a filter in the form of a low-pass, high-pass, and / or band-pass filter, followed by spectral analysis using a Fourier or Hilbert transform. A low-pass filter can be understood as an electronic filter unit that removes high-frequency signal components and retains only low-frequency components. A high-pass filter is a filter mechanism that suppresses low-frequency signal components and allows higher frequency components to pass through. A band-pass filter combines the properties of low-pass and high-pass filters by allowing only a defined frequency range to pass through unchanged. A Fourier transform describes a mathematical method for converting signals from the time domain to the frequency domain, thereby revealing the frequency components of a signal.A Hilbert transform is a mathematical operation for determining the envelope and instantaneous frequency of a signal, which is particularly important for the analysis of non-stationary signals. This inventive use of a combination of filters and spectral analysis enables the targeted separation and analysis of relevant signal components by removing interfering frequency components and extracting characteristic features of the muscle fiber types. This improves data quality, allowing for more precise feature analysis and subsequent classification.
[0080] According to a further embodiment of the invention, the acquisition is performed via an interface of the electronic circuit, which is connected to a transducer for transmitting the high-frequency data. An interface can be understood as an electronic connection unit that enables data exchange between different system components. A transducer is an ultrasonic component that generates high-frequency sound waves and acquires reflected signals for further processing. This use of a direct interface connection, as described in the invention, ensures low-loss and efficient transmission of the ultrasonic data to the data processing device, thereby enabling real-time processing of the signals. This improves the overall performance of the system by realizing fast signal acquisition, processing, and analysis in a compact architecture.
[0081] Further training, functions and advantages of the data processing device result in analogy to the procedure described above.
[0082] The invention further relates to an ultrasound imaging device comprising a transducer and a data processing device, wherein the transducer is connected to the data processing device for signal transmission. In this context, a further embodiment of the invention provides that the transducer is a linear transducer, and the transmission frequency of the ultrasound emitted by the transducer is between 10 and 15 MHz. Signal transmission refers to the transfer of electrical or digitized signals between two units, the connection being a direct physical or wireless coupling between the components of the device.This use of the ultrasound imaging device according to the invention enables continuous and precise acquisition of ultrasound data by ensuring that the signals acquired by the transducer are transmitted to the data processing device without delay and with minimal information loss. This ensures optimal data quality, which allows for improved processing and analysis of the muscle fiber structure.
[0083] Further training, functions and advantages of the ultrasound imaging device result in analogy to the procedure and data processing device described above.
[0084] The invention also relates to a non-volatile, computer-readable storage medium with instructions stored on it which, when executed on a processor, effect the method described above.
[0085] From the data processing device described above for executing the process steps, the ultrasound imaging device with this data processing device, and the non-volatile, computer-readable storage medium containing the instructions that effect the process when executed on a processor, it is evident that the method according to the invention described above is a computer-implemented method, the process steps of which are executed by the data processing device and whose instructions are stored on the computer-readable storage medium. Thus, according to the invention, the method is a computer-implemented method.
[0086] Further developments, functions, and advantages of the computer-readable storage medium arise by analogy to the method, data processing device, and ultrasound imaging device described above. The invention is explained in more detail below with reference to the accompanying drawings and exemplary embodiments. However, it is important to note that these exemplary embodiments do not limit the invention but merely represent embodiments. The features shown can be implemented individually or in combination with other features described in the description and claims.
[0087] The drawings show
[0088] Fig. 1 schematically shows a method for determining the muscle fiber type composition of an individual according to an embodiment of the invention.
[0089] Fig. 2 schematically shows a network model according to an embodiment of the invention, and
[0090] Fig. 3 shows a data processing device according to an embodiment of the invention.
[0091] Figure 1 schematically shows a method for determining the muscle fiber type composition of an individual according to an embodiment of the invention.
[0092] The method comprises a network model 4 with a classifier 1. Network model 4 is an artificial neural network trained to determine muscle fiber type composition. Classifier 1 of network model 4 assigns data acquired via ultrasound imaging of a human muscle to a plurality of different muscle fiber types based on muscle-type-specific characteristics of the muscle data.
[0093] For this purpose, network model 4 was trained using a training dataset that contained data from a majority of people corresponding to the majority of muscle fiber type compositions of the muscle. The majority of muscle fiber types include one fast-twitch and one slow-twitch muscle fiber type.
[0094] The sample consists of six subjects with high speed and three subjects without high speed. The subjects classified as high speed and strength are track and field athletes in the sprint (100m - 400m) and jumping (long jump and high jump) disciplines. Six images were recorded for each subject, in three different ultrasound positions on both the right and left leg, focusing on the calf and thigh muscles. The subjects without high speed and strength are untrained individuals.
[0095] The ultrasound parameters for the training dataset and subsequent data acquisition were selected to enable the recording and processing of muscle data. The penetration depth of the ultrasound signal is set to a range of 12 to 17 millimeters, allowing for highly precise imaging of the relevant muscle layers. The focus is set at 40 millimeters to optimize signal convergence at the depth relevant for muscle fiber type analysis.
[0096] To improve contrast and the visualization of fine tissue structures, the dynamic range is set at 78 decibels, ensuring a balanced ratio between signal amplification and noise reduction. The ultrasound pulse transmission power is zero decibels to guarantee sufficient signal strength while minimizing unwanted artifacts.
[0097] The amplification of the reflected signals is set to 92 percent, resulting in a uniform signal intensity across the entire depth range. The ultrasound frequency used is 12 megahertz, which enables high spatial resolution and is particularly suitable for detailed imaging of muscle fiber structures.
[0098] To optimize the image data, the ultrasound interface was removed to ensure that only relevant image information was included in the analysis. Subsequently, the pixel gray values were normalized to the range between zero and one, guaranteeing uniform scaling of image intensities and preventing differences in brightness or contrast of the original images from causing distortions in the later classification.
[0099] To ensure a sufficient data basis for training the neural network despite a small initial number of participants, each image was divided into four equally sized sections. This division ensures that the available data is used more effectively and that the model learns a greater variability in muscle structures.
[0100] To further increase data diversity and prevent over-adaptation to the training data, targeted data augmentation was performed. The images were transformed through mirroring, rotation, translation, and scaling to ensure the network remains robust against natural variations in image acquisition. Additionally, the images were grayscale inverted to improve the visibility of relevant features, thereby adjusting the contrast properties and enhancing the differentiation of muscle fiber structures. The pre-processed image data was then used to train the neural network.
[0101] The artificial neural network was supervised training with 100 epochs, and the resulting trained artificial neural network, network model 4, as shown in Fig. 2, has 10 layers, which can be divided into an extraction module 2, a pooling module 3, and the classifier 1. The extraction module 2 has three two-dimensional convolution layers that are alternately linked to three pooling layers of the pooling module 3.
[0102] A first two-dimensional convolutional layer serves as the input layer and uses 32 filters with a size of three by three. This layer applies a rectified linear activation function to extract muscle-type-specific features from the data. This is followed by a pooling layer with a pool size of two by two, which reduces the number of features and thus improves the network's computational efficiency. Next comes another two-dimensional convolutional layer with 64 filters of size three by three and a rectified linear activation function to detect increasingly complex patterns from the data. This layer is also followed by a pooling layer, which again reduces the number of features. A third two-dimensional convolutional layer with 128 filters of size three by three and the same activation function further refines feature extraction. A third pooling layer reduces the dimensional data for subsequent processing.
[0103] In the following classifier 1, the resulting muscle-type-specific features are then converted into a vector representation by a transformation layer that transforms the two-dimensional feature maps into a one-dimensional vector. This vector is further processed by a fully connected layer of 128 neurons, which is also equipped with a rectified linear activation function. To improve the generalization of the model and prevent overfitting to the training data, a regulatory dropout layer with a dropout rate of 0.4 is included, which randomly deactivates neurons and thus increases the robustness of the network.
[0104] The final classification is performed by another fully connected layer with two neurons, which uses a softmax activation function. This function allows a probability distribution for the proportions of different muscle fiber types to be output in the form of muscle fiber type compositions.
[0105] In a first step, S1, the ultrasound imaging data of a muscle from an individual are acquired. In a subsequent step, S2, the muscle image data is processed to ensure uniform feature standardization. This processing includes segmentation in step S2a, followed by background removal and filtering of the muscle image data in step S2b. Finally, in step S2c, the image data is normalized. In step S3, the muscle-type-specific features are extracted from the processed image data using extraction module 2 of network model 4 and reduced using pooling module 3. In step S4, the muscle-type-specific features are inputted into classifier 1, which then provides the muscle fiber type composition based on the input of these features.In step S5, the areas from which the muscle-type-specific characteristics were extracted are displayed in the form of a heat map.
[0106] Fig. 3 shows a data processing device 6 according to an embodiment of the invention. The data processing device 6 has an electronic circuit 5 in the form of an integrated circuit. The electronic circuit 5 is configured to perform steps S1 and S2, namely, to acquire data obtained by ultrasound imaging in the form of high-frequency data of a muscle of an individual and to process the high-frequency data of the muscle such that a uniform feature classification is achieved in the high-frequency data. The network model 4 is implemented on a chip 7 in the form of an embedded system and is connected to the electronic circuit 5 for signal transmission in order to perform the steps of extracting muscle-type-specific features from the processed data and providing the muscle fiber type composition. In this case, the chip 7 is a microprocessor.
[0107] 1 classifier
[0108] 2 Extraction module
[0109] 3 Pooling module
[0110] 4 Network model
[0111] 5 electronic circuit
[0112] 6 Data processing device
[0113] 7 Chip
Claims
Patent claims 1. A method for determining the muscle fiber type composition of an individual using a classifier (1), wherein the classifier (1) assigns muscle fiber types of a multiple of different muscle fiber types proportionally to muscle fiber types acquired by ultrasound imaging of a human muscle based on muscle-type-specific characteristics of the muscle data, comprising the following process steps: 51) Acquiring data of a muscle of the individual acquired by ultrasound imaging, 52) Processing the muscle data so that there is a uniform feature standardization in the data, 53) Extracting muscle-type-specific features from the processed data, taking into account the different muscle fiber types exhibiting different ultrasound reflectivities based on the muscle-type-specific features, and 54) Inputting the muscle-type-specific features into the classifier (1), wherein the classifier (1) provides a composition of the proportions of different muscle fiber types in the form of the muscle fiber type composition in response to the input of the muscle-type-specific features.
2. The method of claim 1, wherein the data are image data in the form of ultrasound images and / or radio frequency data.
3. Method according to claim 1 or 2, wherein in the step of acquiring the data of the individual's muscle, instead: 51) Acquisition of data acquired by ultrasound imaging of a muscle of an individual, wherein the data were acquired by aligning a transducer of an ultrasound imaging device longitudinally to the muscle fibers of the muscle in a time-dependent manner during a contraction of the muscle, so that the data show a time-dependent course of the muscle fibers in the form of an image sequence, and the subsequent steps are instead: 52) Processing the image sequence so that there is a uniform feature standardization in the image sequence, and S3) Extracting muscle-type-specific features from the processed image sequence, wherein a muscle-type-specific feature is a contraction velocity of the muscle.
4. Method according to one of the preceding claims, wherein the classifier is configured to determine the muscle fiber type composition with spatial resolution, and in the step of acquiring the data of the individual's muscle instead: Sl) Acquisition of data acquired by ultrasound imaging of a muscle of an individual, wherein the data were acquired by aligning a transducer of an ultrasound imaging device transversely to the muscle fibers of the muscle, so that the data perpendicular to a skin surface of the individual show the depth of the muscle, and the step of inputting the muscle type-specific features instead: S4) Inputting the muscle type-specific features into the classifier (1), wherein the classifier (1) provides the muscle fiber type composition in a region of the muscle in response to the input of the muscle type-specific features.
5. Method according to any of the preceding claims, wherein the muscle type-specific features are extracted using a gray value matrix and / or using statistical features and / or using edge or flank detection and / or Fourier or wavelet transformation and / or principal component analysis.
6. A method according to any of the preceding claims, wherein the method in the step of processing the muscle data, such that a uniform feature standardization is present in the data, comprises the following process steps: S2a) Segmenting the muscle data, S2b) Removing a background and filtering the muscle data, and S2c) Normalizing the muscle data.
7. Method according to one of the preceding claims, wherein the classifier (1) provides a muscle fiber type space in which a plurality of classes for different muscle fiber type compositions are formed, and the classifier (1) assigns the data to one of the classes based on the muscle type-specific features extracted from the data, and the muscle fiber type composition is determined based on distances of the muscle data to centers of the classes.
8. Method according to one of the preceding claims comprising an artificial neural network trained to determine the muscle fiber type composition in the form of a network model (4), wherein the network model (4) has the classifier (1), the network model (4) was trained using a training dataset, and the training dataset contains data from a plurality of people corresponding to a plurality of different muscle fiber type compositions of the muscle, and the network model (4) provides the muscle fiber type composition with the classifier (1) in response to the input of the muscle type-specific features.
9. Method according to the preceding claim, wherein the training data set was enlarged by mirroring and / or rotating and / or translating and / or scaling and / or splitting the data of the majority of people.
10. Method according to one of claims 8 or 9, wherein the network model (4) comprises an extraction module (2) and a pooling module (3), wherein the extraction module (2) extracts the muscle type-specific characteristics in the step of extracting muscle type-specific characteristics from the processed data, the extraction module (2) is linked to the pooling module (3), wherein the pooling module (3) reduces the muscle-type-specific features, the pooling module (3) is linked to the classifier (1), and The classifier (1) analyzes the reduced muscle type-specific features and provides the muscle fiber type composition based on the analysis.
11. A method according to any one of claims 8 to 10, wherein the extraction module (2) comprises a plurality of two- or three-dimensional convolution layers, the pooling module (3) comprises a number of pooling layers corresponding to a number of convolution layers, and the convolution layers are alternately linked with the pooling layers.
12. A data processing device (6) for carrying out the method steps according to any one of the preceding claims.
13. Data processing device (6) according to the preceding claim, comprising an electronic circuit (5), wherein the electronic circuit is configured to acquire data acquired by means of ultrasound imaging in the form of high-frequency data of a muscle of an individual and to process the high-frequency data of the muscle, a chip (7) connected to the electronic circuit (5) for signal transmission and comprising the classifier (1), wherein the chip (7) is designed to extract the muscle type-specific characteristics from the processed high-frequency data, and The classifier (1) provides the muscle fiber type composition in response to the input of the muscle type-specific characteristics.
14. Ultrasound imaging device comprising a transducer and the data processing device (6) according to one of the two preceding claims, wherein the transducer is connected to the data processing device (6) for signal transmission.
15. Non-volatile, computer-readable storage medium containing instructions stored thereon which, when executed on a processor, effect a method according to any one of claims 1 to 11.
