Techniques for determining indicators of a medical condition

Abstract

A medical data processing technique for determining an indicator of a medical condition is disclosed. A method implementation of the technique includes selecting (202) at least one model from a plurality of models based on at least one characteristic associated with medical data of a test instance, each of the plurality of models being generated by a learning algorithm and configured to provide a model-specific indicator of the medical condition, determining (204) a respective model-specific indicator using each of the at least one selected model, and determining (206) the indicator of the medical condition based on the model-specific indicator.

Description

[Technical Field]

[0001] The present disclosure relates generally to the field of medical data processing. In particular, techniques are presented for enabling the determination of indicators of medical conditions. The techniques may be embodied in the form of methods, computer programs, and apparatuses. [Background technology]

[0002] Computer-aided technologies for medical diagnosis, including machine learning techniques, have advanced in recent years and are considered powerful tools for assisting physicians in medical diagnosis. Research work supports the assertion that such technologies have the potential to significantly improve physician or radiologist workflows, either in terms of diagnostic quality (e.g., improving the ability to spot malignant pulmonary nodules in chest computed tomography (CT) images) or time efficiency (e.g., reducing time to diagnosis from minutes to seconds). Aside from medical imaging, workflows in other diagnostic pillars also hold the potential to be improved by these technologies, for example, in electroencephalography (EGG) (e.g., detecting signs of epilepsy in EEG signals) or laboratory test results (e.g., predicting hematological diseases from blood test results).

[0003] Existing solutions have been shown to produce high prediction performance. However, they suffer from several drawbacks. For example, some known approaches exhibit poor generalization. In particular, these approaches may be practically unusable due to their inapplicability to data types different from those used for training, which may occur, for example, when the recording method (e.g., CT images, magnetic resonance (MR) images, two-dimensional X-ray images, or EEG signals) or data quality (e.g., image resolution or sample rate of EEG signals) is different. Furthermore, some currently known approaches exhibit low robustness, such that even a small change in the input data can result in large changes in the determined predictions. Furthermore, some known approaches may be unsafe to use due to their inability to reflect the uncertainty of the determined predictions. Last but not least, deep learning models are generally not well explainable because their outputs are generated by trained black-box functions and therefore may be considered insufficiently traceable by humans. This can reduce user confidence and complicate bug fixing. Summary of the Invention

[0004] Therefore, there is a need for technological implementations that can provide reliable and versatile determination of indicators of medical conditions.

[0005] According to a first aspect, a medical data processing method for determining an indicator of a medical condition is provided. The method includes selecting at least one model from a plurality of models based on at least one characteristic associated with medical data of a test instance. Each of the plurality of models is generated by a learning algorithm and configured to provide a model-specific indicator of the medical condition based on the medical data. The method further includes determining a respective model-specific indicator of the medical condition based on the medical data using each of the at least one selected model. The method also includes determining the indicator of the medical condition based on the model-specific indicator.

[0006] At least one model may be selected from among a plurality of models based on at least one characteristic, where the characteristic may be related to the test instance. Thus, the selection may depend on the test instance at hand. The selection may also be referred to as dynamic selection. Such dynamic selection may ensure robust determination of an indicator of a medical condition based on at least one characteristic. By providing different models suitable for different characteristics of the medical data, the most suitable model may be selected depending on the medical data at hand. For example, if at least one characteristic is the image modality of a medical image included in the test instance medical data, only models capable of determining an indicator based on such image modality may be selected. The at least one selected model may be generated by a learning algorithm using training data of the same image modality as the test instance medical data. This may improve the reliability of the determined indicator.

[0007] Two or more models can be selected from the plurality of models. Each of the selected models can then provide a respective model-specific index, and a (e.g., final) index of the medical condition can be determined based on all these model-specific indexes. For this reason, a (e.g., group of) selected models, each generated by a (e.g., different) learning algorithm, can be referred to as an ensemble. If the selected models are generated by a machine learning algorithm, the (e.g., group of) selected models can be referred to as a machine learning ensemble. The outputs of different models can advantageously be aggregated or combined to determine an index. Because different models may produce different model-specific indexes based on the same medical data, such aggregation or combination can increase the overall reliability of the determined index.

[0008] A test instance may be, for example, a dataset related to a patient. The dataset may be medical data and, optionally, data (e.g., metadata) describing or defining at least one characteristic associated with the medical data of the test instance. A patient may be referred to as a patient to whom a medical diagnosis or an indication of a medical condition is to be assigned. Medical data, in other words, relates to a patient. The medical data may relate to a patient in that it describes a medical characteristic of the patient. The medical data may include data related to (e.g., describing or representing) a measurement of at least one characteristic of the patient's body, where the measurement may be a medical measurement. The medical data may relate to (e.g., describing or representing) a condition of the patient's body, where the condition may be an anatomical, physical, or physiological condition. The medical data may include data describing EEG measurements, data describing blood test results, data describing medical images, etc. The medical data may be omics data, such as genomics, proteomics, glycomics, immunonomics, brainomics, etc. The medical data may include or be a medical image (e.g., of a patient). The medical image may be collected by medical image acquisition (of at least a portion of the patient's body), and may be, for example, a CT image, an MR image, an ultrasound image, or a camera image. The step of collecting medical data may be, but is not necessarily, part of the methods disclosed herein. The medical image may relate to a patient in that it includes or consists of at least one medical data of the patient.

[0009] A medical condition can describe or represent a characteristic or distinctive feature (e.g., anatomical, physical, or physiological) specific to the patient to whom the test instance pertains. A medical condition or indicator of a medical condition can be a predicted or determined (e.g., class or higher-level) label for the medical data of the test instance. A medical condition or (e.g., class or higher-level) label can describe that the patient is healthy, that the patient is unhealthy, that the patient has a certain (e.g., predetermined) symptom, that the patient exhibits a medical abnormality (e.g., compared to healthy patients), etc.

[0010] The learning algorithm may be an algorithm that learns (or "optimizes") using labeled or unlabeled training data as input, where the model may be the output of the learning algorithm. The learning algorithm may be an algorithm used to determine (or "optimize") parameters in or to be used by a model (e.g., one model of a plurality of models), where labeled or unlabeled training data may be used as input to learn or determine (e.g., optimal) parameters. The learning algorithm may define a particular sequence of method steps and may be used, for example, to optimize parameters of a statistical model. In another variation, the learning algorithm may be a machine learning algorithm that optimizes parameters of a machine learning model. At least one of the plurality of models may be, for example, an adversarial autoencoder or a neural network.

[0011] The characteristics associated with the medical data of the test instance may be characteristics described by metadata of the medical data (e.g., in a header of a medical image) or characteristics extracted from (content contained in) the medical data (e.g., a medical image) of the test instance. The characteristics associated with the medical data of the test instance may be specific to the test instance and / or the medical data of the test instance. For example, the at least one characteristic associated with the medical data may include or be a feature of the medical image (e.g., slice thickness, image resolution, image contrast, radiomics features, image modality, etc.). The at least one characteristic associated with the medical data may include or be a characteristic of the patient to which the medical data relates (e.g., age, gender, ethnicity, etc.).

[0012] The step of selecting at least one model may be performed, for example, by a selector. The selector may include or be configured with a selector model. The selector may represent or define some steps of the method described herein. The step of selecting at least one model may be performed before (or "prior to") determining the model-specific indicators. The model-specific indicators may be determined only by the selected at least one model. In other words, the step of determining each model-specific indicator of the medical condition based on the medical data may be performed using only each of the at least one selected model. This can reduce the amount of computational time and resources required to determine the indicators of the medical condition.

[0013] In one variant, the selection step comprises comparing, for each of the plurality of models individually, at least one characteristic associated with the medical data of the test instance with at least one characteristic associated with the training data used to generate the respective model, thereby ensuring that (only) the most suitable models are selected from the plurality of models for (for example) determining an indicator of a medical condition, resulting in a (e.g. sufficiently suitable or optimal) ensemble of models dynamically selected based on metadata of (e.g. medical data of) the test instance depending on the use case.

[0014] The training data may include or consist of medical data for multiple different patients. The training data may include or consist of multiple medical images (e.g., of different patients), also referred to as training images. The training images may be collected by medical image acquisition, such as CT images, MR images, ultrasound images, camera images, etc. The training data may include or consist of different types of content, such as EEG signals, blood test results, genomics, etc. The characteristics associated with the training data may be characteristics described by metadata (e.g., in a header or label) of the training data or characteristics extracted from (e.g., content contained within) the training data (e.g., training images). The characteristics associated with the training data may be specific to the training data. For example, at least one characteristic associated with the training data may include or be at least one feature of the training images (e.g., slice thickness, image resolution, image contrast, radiomics features, image modality, etc.), which may be included in or correspond to the training data. The at least one feature associated with the training data may include or be a characteristic of one or more patients to which the training data pertains (e.g., age, gender, ethnicity, etc.) In one variation, the at least one feature associated with the medical data of the test instances and the at least one feature associated with the training data used to generate the individualized model may be of the same feature type.

[0015] The indicator of the medical condition can further be determined based on at least one attribute selected from the results of the comparison step, the empirical performance of each of the plurality of models, and the degree of explainability of each of the plurality of models. The at least one attribute can be specific to each model. This can enable one model to be prioritized over another model according to the model-specific attribute in order to determine the indicator of the medical condition with higher reliability. For example, results obtained by a model with a relatively low suitability for determining the model-specific indicator can be given less weight than a model with a higher suitability for determining the model-specific indicator.

[0016] The result of the comparison may be a suitability measure (e.g., expressed as a numerical or binary value) of each of the plurality of models for determining a model-specific index for the medical data of the test instance. The suitability measure may be obtained from a database storing a plurality of suitability measures associated with one or more (e.g., possible or range) characteristics of the medical data. The empirical performance of each model among the plurality of models may be determined by using predetermined, preferably labeled, medical data to determine each model-specific index, and measuring the performance of each model. The performance may be temporal performance (e.g., the time it takes for the model to provide the model-specific index using predetermined, limited computational resources), resampling performance, and / or confidence performance (e.g., the probability or degree of false positive detection, the probability or degree of false negative detection, etc.). The empirical performance may be obtained from a database storing a plurality of empirical performance measures associated with one or more models.

[0017] The explainability of a model may be the degree to which the results output by the model are understandable or traceable by humans. The higher the explainability of a model, the easier it is for a user to understand how the model-specific indicators are determined using the model. The explainability of a model may be a value (e.g., a numerical or binary value) associated with the model that can be determined or defined manually. For example, a low explainability may be determined for a model that relies on one or more neural networks because the model's output may be considered to be generated by a (trained) black-box function, while a higher explainability may be determined for a model whose output results are more traceable than those of neural networks. The explainability may be predetermined for one or more, preferably all, of the multiple models. The explainability may be obtained from a database, for example, a database storing multiple explainability values ​​associated with one or more of the multiple models.

[0018] The indicator of the medical condition can be determined based on at least the model-specific indicators by an aggregator. The aggregator can represent or define some steps of the methods described herein. The step of determining the indicator can occur after (or "subsequent to") the step of selecting at least one model and determining the model-specific indicator. As noted above, this can avoid determining model-specific indicators that are not used to determine the (e.g., final) indicator of the medical condition, thereby reducing the amount of computational time and resources required to determine the indicator.

[0019] The (e.g., final) indicator of the medical condition can be determined (e.g., by an aggregator) as an average of the model-specific indicators, e.g., a weighted average where the weights are based on at least one attribute, in particular empirical performance. As pointed out above, this allows prioritizing models according to at least one attribute, thereby improving the reliability of the determined indicator of the medical condition. The indicator can be determined (e.g., by an aggregator) as a majority vote of the discrete (e.g., binary) model-specific indicators. Depending on the use case, for example, depending on the medical data at hand (e.g., its type) and / or depending on the medical condition to be determined, different votes of the model-specific indicators can be used to determine the indicator. For example, if the medical condition to be determined is whether a patient is healthy or whether an abnormality exists in the patient's medical data, it may be useful to determine that an abnormality exists if the model-specific indicator of at least one of the selected models so indicates. This may increase the safety of use of the method. The index can be determined (e.g., by an aggregator) based on a subset of the model-specific indexes, where the subset can be determined based on at least one attribute, particularly based on the results of the comparison. For example, a user may not trust the predictions of a neural network and therefore want to ignore the model-specific indexes determined by the neural network. In other words, model-specific indexes of models with low explainability can be ignored for determining an index of a medical condition. Alternatively, such models may not be selected (e.g., by a selector) in the step of selecting at least one model. As another example, individual models may determine erroneous outliers outside a certain range as model-specific indexes, while several selected models may all provide model-specific indexes within the certain range. Model-specific indexes outside the certain range can be ignored (e.g., by an aggregator) for determining an index of a medical characteristic. As mentioned above, this can enable a more reliable determination of an index of a medical condition.

[0020] The (e.g., final) indicator of the medical condition may be determined using an aggregation model, as noted above. For example, the aggregator may define, include, or be composed of the aggregation model. The aggregation model may be generated by an aggregation model learning algorithm. The aggregation model learning algorithm may be a machine learning algorithm. A plurality of model-specific indicators determined by a plurality of models based on a plurality of test instances may be used as inputs for training the aggregation model. Optionally, at least one feature associated with the aggregation training data may be used as inputs for training the aggregation model, along with the plurality of model-specific indicators. The at least one feature associated with the aggregation training data may be of the same type as the at least one feature associated with the medical data and / or the at least one feature associated with the training data. The aggregation model learning algorithm may be an algorithm for training the aggregation model, such as, for example, a backpropagation algorithm when the aggregation model is a neural network. The use of aggregation models can improve the reliability of the determined indicator of a medical condition, especially when more information than just multiple model-specific indicators (i.e., at least one characteristic) is considered.

[0021] A supervised learning algorithm may require labeled training data. The label may indicate whether a certain (e.g., predetermined) medical condition or set of medical conditions applies to the training data. The labeled training data can then be used to train the supervised learning algorithm so that it can predict the label as a model-specific indicator when applied to new data that has not been used for training. At least one of the models included in the plurality of models can be generated by the supervised learning algorithm, for example, using labeled training data (e.g., of at least one healthy patient and at least one unhealthy patient with a disease).

[0022] In particular, in medical diagnosis tasks, the structure of the labels to be predicted may be hierarchical. There may be a higher-level label (also called a class label) for a class (e.g., "normal" corresponding to "healthy" or "abnormal" corresponding to "unhealthy"). Some of the classes (e.g., the "abnormal" class) may have multiple subclasses (e.g., different subclasses for different diseases such as "lung cancer," "prostate cancer," and "melanoma") that may be associated with lower-level labels (also called subclass labels).

[0023] If a model is generated by a supervised learning algorithm using labeled training data, it may only be able to provide (e.g., determine) a model-specific indicator of whether a test instance is associated with one of the labels on which the model was trained. In this case, determining a model-specific indicator of a medical condition associated with the medical data of a test instance (e.g., whether the medical condition is “healthy” or “abnormal”) may be possible only if the training data represents a balanced amount of all possible diseases that could potentially be detected. In other words, labeled data for each possible subclass may need to be available to train a supervised learning algorithm to enable correct prediction of the label of a subclass (e.g., “lung cancer”) or each (e.g., higher-level) class (e.g., “abnormal”). In particular, for diseases with very low prevalence, such training data may be difficult to find or unavailable. As a result, using a model generated by a supervised learning algorithm to determine a model-specific indicator of a medical condition associated with the medical data of a test instance (e.g., whether the medical condition is “healthy” or “abnormal”) may be difficult in some cases.

[0024] On the other hand, a reliable indicator of a medical condition of “healthy” or “abnormal” may enable a physician to prioritize and appropriately allocate time for cases even before a consultation. Thus, in one particular variant, for example, unlabeled training data (e.g., only) (e.g., of healthy patients) may be used by a learning algorithm to generate at least one of the models (e.g., at least one selected model) included in the plurality of models. In a practical use case, medical data taken from a big register study, in which it can be assumed that the majority of patients are generally healthy, may be used as the unlabeled training data. In this case, the unlabeled training data may also be referred to as “weakly labeled” training data due to the assumption that the majority of the patients are healthy. Instead of the unlabeled training data, training data having the same (e.g., higher-level) label may be used to generate at least one of the models (e.g., at least one selected model) included in the plurality of models. In a particular variant, the learning algorithm may be an unsupervised (e.g., machine) learning algorithm. At least one of the models included in the plurality of models may be configured to provide (e.g., determine) anomaly detection as a model-specific indicator of a medical condition. In particular, at least one of the models included in the plurality of models may be generated by an unsupervised learning algorithm using (e.g., unlabeled) training data (only) of (e.g.,) healthy patients, and may optionally be configured to provide (or determine) anomaly detection as a model-specific indicator of a medical condition. The anomaly detection may, for example, correspond to out-of-distribution detection. Using anomaly detection as out-of-distribution detection for determining a model-specific indicator allows for a reliable determination that a patient is not healthy, even if training data representative of the patient's disease was not used to generate the at least one model. Furthermore, using unlabeled training data of healthy patients allows for easy acquisition of a large training dataset.

[0025] The model-specific index of the medical condition and / or the index of the medical condition may include a numerical value describing at least one result selected from the probabilities of abnormalities of different portions of the medical image included in the medical data and the probability of abnormality of the overall medical data. Optionally, the numerical value may be derived from the probabilities of abnormalities for different portions of the medical image. The model-specific index and / or index may include a probability of a certain medical condition. This approach may be safe to use because it may have the ability to reflect uncertainty in the determined model-specific index and / or determined index.

[0026] In a first variation of the present disclosure, the determined model-specific indicator and / or the determined (e.g., final) indicator of a medical condition may include or be, for example, a probability that a predetermined medical condition exists in the test instance. The probability may be a numerical or binary value. The (e.g., predetermined and / or model-specific) medical condition may be that the patient is healthy, that the patient is not healthy, that the patient has a certain (e.g., predetermined) symptom, that the patient exhibits a medical abnormality (e.g., compared to a healthy patient), etc.

[0027] In a second variant, the determined model-specific indicator of the medical condition and / or the determined indicator may include or be an identification of a portion of the medical data associated with (e.g. a predetermined) medical condition. For example, the model-specific indicator of the medical condition and / or the indicator of the medical condition may include or be an identification of a portion of a medical image included in the medical data, which portion may be associated with (e.g. a predetermined) medical condition (e.g. a portion of an image associated with cancerous tissue). The portion of an image (e.g. a medical image) may be, for example, a pixel, a voxel, an area or a volume.

[0028] In a variant combining the first and second variants, the determined model-specific indicator and / or determined indicator of a medical condition may include or be a probability that a portion of medical data (e.g., a portion of a medical image) is associated with a (e.g., predetermined) medical condition.

[0029] The method may further include determining (e.g., by the selector and / or aggregator) that a reliable determination of the indicator is impossible if at least one characteristic associated with the medical data of the test instance does not indicate the suitability of at least one model (e.g., suitability for determining a model-specific indicator for the medical data of the test instance). The step of determining that a reliable determination of the indicator is impossible may be based on the result of the comparison. For example, if the result of the comparison is a suitability, it may be determined that a reliable determination of the indicator is impossible if the suitability does not meet a predetermined criterion (e.g., is lower than a predetermined threshold). In one variant, it may be determined that a reliable determination of the indicator is impossible if no suitable model can be selected. If a reliable determination of the indicator is determined to be impossible, the method may not determine the model-specific indicator and / or the indicator, and / or the method may include triggering the output of a notification on an output device informing a user that a reliable determination of the indicator is impossible. In this way, the determination of an unreliable indicator of the medical condition can be prevented.

[0030] If it is determined that a reliable determination of an indicator of the medical condition is possible, or if an indicator of the medical condition has already been determined, the method may comprise the step of triggering the output of a notification on an output device informing the user about the indicator or that a reliable determination of the indicator is possible. The notification and / or the indicator may be visualized on the display accordingly.

[0031] The method may further include determining (or "making") a medical diagnosis based at least on the determined index. The determination of the medical diagnosis may further be based on a value described by at least a portion of the medical data referred to or described by the index. The determination of the medical diagnosis may include comparing the determined index with a medical diagnosis set correlated with different indexes, and selecting a (e.g., matching or most suitable) medical diagnosis from the diagnosis set, for example, based on the comparison step. The medical diagnosis may be, for example, that the patient has a brain tumor, an area or volume of cancerous tissue in their body, that the patient is healthy, that the patient has amyotrophic lateral sclerosis (ALS), etc.

[0032] In one implementation, at least one of the models included in the plurality of models, hereinafter also referred to as density model, can correlate portions of a medical image and portions of a reference image included in or represented by the medical data. For example, a (e.g., predetermined) first registration between the medical image and the reference image can be applied to relate both to a common reference frame. The first registration can include a transformation of the coordinate system of the medical image (e.g., a portion thereof) to the coordinate system of the reference image (e.g., a portion thereof), or vice versa. The first registration can include or consist of a first image transformation. The first registration can be applied to one or more portions of the medical image and only the reference image. Different first registrations can be applied to different portions of the medical image and the reference image. In one variant, the medical image can be aligned to the reference image to correlate portions of the medical image with portions of the reference image. The alignment can be performed, for example, based on an image fusion algorithm. The reference image may be a medical image, an image of a standard anatomical model, an image obtained by averaging multiple patient images registered to a common reference frame, or the like.

[0033] The density model may further compare (e.g., provide a comparison of) image values ​​(e.g., at least one image value or all image values) of at least one portion of a medical image contained in or represented by the medical data with information associated with a correlated portion of a reference image to obtain a model-specific indicator of a medical condition.

[0034] The information to which image values ​​are compared may be generated or determined by a learning algorithm that defines the sequence of method steps. For example, the information to which image values ​​are compared may be generated or determined by aligning a plurality of training images with a base image and correlating portions of each training image of the training images with portions of the base image (e.g., on a processor or device different from the processor or device that performs the method of the first aspect and / or before performing the method of the first aspect). The alignment may be performed, for example, using an image fusion or image alignment algorithm. Alternatively, or additionally, the alignment step may include or consist of applying a (e.g., predetermined) second registration between each (e.g., a portion thereof) of the plurality of training images and the base image (e.g., a portion thereof) to relate each training image (e.g., a portion thereof) of the plurality of training images and the base image (e.g., a portion thereof) to a common reference frame. The second registration may include a transformation of the coordinate system of one medical image (e.g., a portion thereof) into the coordinate system of the base image (e.g., a portion thereof), or vice versa. Thus, the second registration may include or consist of a second (image) transformation. The second registration may be applied to only one or more portions of each training image and one or more portions of the reference image. Different second registrations may be applied to different portions of each training image and reference image.

[0035] The information to which image values ​​are compared may further be generated or determined by determining image values ​​of at least one portion of each training image of a plurality of training images correlated with a portion of the base image, or may have been generated or determined (e.g., on a different processor and / or prior to performing the method of the first aspect), where the portion of the base image may be assigned (e.g., correlated, mapped, or aligned) to the correlated portion of the reference image using a (e.g., third) predetermined transformation. The portion of the base image may be assigned to the correlated portion of the reference image using a predetermined third registration. The third registration may include a transformation that transforms the coordinate system of the base image (e.g., portion thereof) to the coordinate system of the reference image (e.g., portion thereof), or vice versa. Thus, the third registration may include or consist of a third (image) transformation. The predetermined third transformation or registration may relate both the reference image (e.g., correlated portion thereof) and the base image (e.g., portion thereof) to a common frame of reference.

[0036] The image value of the at least one portion may be the value of a single pixel or voxel in each of the respective training images (e.g., density value, Hounsfield value, color value, saturation value, etc.). The image value of the at least one portion may be the average, maximum, or minimum of pixel or voxel values ​​of pixels or voxels included in the at least one portion in each of the respective training images.

[0037] The information may be a collection of pixel or voxel values ​​(eg, average, maximum, or minimum) of corresponding portions of multiple training images correlated with portions of the base image.

[0038] The information to which the image values ​​are compared may further be generated or determined by determining information based on determined image values ​​of at least one portion of each of the plurality of training images, or may be generated or determined (e.g., on a different processor and / or before performing the method of the first aspect). The information may, for example, include, be, or consist of a statistical distribution function of the image values ​​of at least one portion of the plurality of training images. The statistical distribution function may be determined as a best fit of a predetermined distribution type (e.g., Gaussian distribution, kernel density estimation, multinomial distribution, etc.) to the image values ​​(e.g., a set of pixel or voxel values) of at least one portion of the plurality of training images. Experiments have shown that using information including a statistical distribution function, the density model can result in a highly reliable determination of the model-specific indicator.

[0039] Alternatively, or additionally, the information may include, be, or consist of an average image value of at least one portion of all of the plurality of training images, e.g., an average of a set of pixel or voxel values. The information may (e.g., further) include, be, or consist of an average deviation of the image values ​​of at least one portion of all of the plurality of training images from the average image value, e.g., an average deviation of each pixel or voxel value of the set of pixels or voxel values ​​from the average of the set of pixel or voxel values. Experiments have shown that using information including the average image value and the average deviation, a density model can result in a reliable determination of model-specific indicators. Similarly, such density models can be generated quickly with limited computational resources.

[0040] According to a second aspect, there is provided a medical data processing method for determining an indicator of a medical condition. The method includes correlating portions of medical data included in the medical data with portions of a reference image. The method further includes comparing image values ​​of at least one portion of the medical image with information associated with the correlated portion of the reference image to obtain the indicator of the medical condition. For example, the information may have been generated or determined, or may have been generated or determined (e.g., on a processor different from and / or at a different time than the processor performing the method of the second aspect) by aligning a plurality of training images with a base image and correlating a portion of each training image of the training images with a portion of the base image; determining an image value of the at least one portion of each training image of the plurality of training images correlated with a portion of the base image, wherein the portion of the base image is assigned to the correlated portion of the reference image using a predetermined transformation; and determining information based on the determined image value of the at least one portion of each training image of the plurality of training images, wherein the information is a statistical distribution function of image values ​​of the at least one portion of the plurality of training images.

[0041] All details and variations of the features of the first aspect, indicated with the same terminology as used for the second aspect, may also be applied to the method of the second aspect, where the indices of the second aspect may correspond to the model-specific indices of the first aspect. The method of the second aspect may also include determining a medical diagnosis based on the determined indices. The method of the first aspect may be combined with the method of the second aspect, and vice versa. The method of the second aspect may be used to determine each model-specific indices of the first aspect. That is, the method of the second aspect may correspond to a method performed by one of the models of the method of the first aspect.

[0042] The method of the first aspect and / or the method of the second aspect may include a step of collecting medical data. The step of collecting medical data may simply consist of collecting data from a storage medium. The step of collecting medical data may not be performed on a human or animal body. In particular, none of the steps included in the method of the first aspect and / or the method of the second aspect may require interaction with a human or animal body. None of the steps included in the method of the first aspect and / or the method of the second aspect may require the presence of a patient's body.

[0043] According to a third aspect, an apparatus is provided. The apparatus includes at least one processor and at least one memory, the at least one memory storing instructions executable by the at least one processor such that the apparatus is operable to perform the method of the first aspect and / or the method of the second aspect. The apparatus may further include an output device and / or an interface for receiving, acquiring, transmitting, or sending data or information. The method of the first aspect and / or the method of the second aspect may be a computer-implemented method. Each step of the method of the first aspect and / or the method of the second aspect may be performed by at least one processor of the apparatus.

[0044] According to a fourth aspect, a computer program is provided. The computer program includes program code portions for performing the method of the first aspect and / or the method of the second aspect when the computer program product is executed on one or more processors. The computer program product may be stored on one or more (e.g., non-transitory) computer-readable recording media. The computer program product may be stored in a carrier such as an electronic signal, an optical signal, or a data stream.

[0045] According to a fifth aspect, there is provided one or more computer-readable storage media storing the computer program product of the fourth aspect. The one or more computer-readable storage media may be non-transitory storage media.

[0046] Further details and advantages of the techniques presented herein are explained in connection with the exemplary implementations shown in the figures. [Brief explanation of the drawings]

[0047] [Figure 1] FIG. 1 shows an exemplary configuration of an apparatus according to the present disclosure. [Figure 2] FIG. 2 illustrates a method for determining indicators of a medical condition that can be performed by a device according to the present disclosure. [Figure 3] FIG. 3 illustrates a schematic diagram of an exemplary method for generating a model that can be performed by an apparatus according to the present disclosure. [Figure 4] FIG. 4 illustrates a schematic diagram of an exemplary method for determining a model-specific indicator of a medical condition that can be performed by an apparatus according to the present disclosure. [Figure 5] FIG. 5 illustrates a schematic diagram of an exemplary method for generating multiple models that can be performed by an apparatus according to the present disclosure. [Figure 6] FIG. 6 illustrates a schematic diagram of an exemplary method for determining an indicator of a medical condition that can be performed by an apparatus according to the present disclosure. [Figure 7] FIG. 7 shows a schematic diagram of an exemplary data label. [Figure 8] FIG. 8 shows a schematic representation of exemplary prevalence rates of different diseases. [Figure 9] FIG. 9 illustrates a method that can be performed by an apparatus according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0048] In the following description, for purposes of explanation and not limitation, specific details are set forth in order to provide a thorough understanding of the present disclosure. It will be apparent to those skilled in the art that the present disclosure may be practiced in other implementations that depart from these specific details.

[0049] Those skilled in the art will further recognize that the steps, services, and functions described herein below can be implemented using discrete hardware, using software working in conjunction with a programmed microprocessor or general-purpose computer, using one or more application-specific integrated circuits (ASICs), and / or using one or more digital signal processors (DSPs). Similarly, where the present disclosure is described in terms of methods, it will also be recognized that the present disclosure can likewise be embodied in one or more processors and one or more memories coupled to the one or more processors, where the one or more memories are encoded with one or more programs that, when executed by the one or more processors, perform the steps, services, and functions disclosed herein.

[0050] FIG. 1 illustrates an exemplary configuration of an apparatus 100 according to the present disclosure. The apparatus 100 includes a processor 102, a memory 104, and optionally an interface 106. The processor 102 is coupled to the memory 104 and optionally the interface 106. The interface 106 is configured to acquire, receive, send, or transmit data from or to an external unit, such as a data storage unit, a server, a user input unit, or an output unit such as a display or speaker. The interface 106 may be configured to send a trigger signal to the output unit to trigger the output of an audio and / or optical notification (message) for a user. The memory 104 is configured to store a program including instructions that, when executed by the processor 102, cause the processor 102 to perform the methods described herein. The program may be stored on a storage medium.

[0051] 2 illustrates a method that can be performed by the apparatus 100. The method of FIG. 2 may correspond to the method of the first aspect described above. The method includes a step 202 of selecting (e.g., by or using a selector described in the present invention) at least one model from a plurality of models based on at least one characteristic associated with medical data of a test instance, where each of the plurality of models is generated by a learning algorithm and configured to provide (e.g., determine) a model-specific indication of a medical condition based on the medical data. Each model can be referred to as a base learner model.

[0052] The method further includes determining 204 a respective model-specific indicator of the medical condition based on the medical data using (eg, only) each of the at least one selected model.

[0053] The method also includes determining 206 (by or using an aggregator as described herein) a (eg, final) indicator of the medical condition based on the model-specific indicator.

[0054] Figure 3 schematically illustrates an exemplary method for generating a base learner model according to the present disclosure. The method illustrated in Figure 3 may be performed by device 100 or by a different device (not shown) that includes a processor and a memory that stores instructions that, when executed by the processor, cause the processor to perform the method illustrated in Figure 3. The method illustrated in Figure 3 can be used to generate at least one of the models used in the method of Figure 2.

[0055] Training data, such as medical images of multiple patients, can be used by a learning algorithm, also referred to below as a base learner, to generate one of the base learner models. A learner configuration (“config”) file can define the type and / or structure of the base learner model. The learner configuration file can include at least one hyperparameter defined in or used by the (e.g., untrained) base learner model. The learner configuration file can define the learning algorithm used to generate the base learner model and, optionally, the hyperparameters defined in or used by the learning algorithm. Some preferred examples of base learners and base learner models are detailed below in connection with FIGS. 7-9. In the method shown in FIG. 3, the base learner model can be generated based on the training data and the learner configuration file. Thus, the training data can be used as input “I,” the learner configuration file can be used as input “I,” and the base learner model can be the determined output “O.” The base learner model can be considered the output of the base learner.

[0056] FIG. 4 schematically illustrates an exemplary method for determining a model-specific indicator for a medical condition according to the present disclosure. The method illustrated in FIG. 4 is performed by apparatus 100 and may be part of the method illustrated in FIG. 2, specifically, part of step 204. In the method illustrated in FIG. 4, the model-specific indicator for the medical condition may be determined based on a test instance, for example, based on medical data of the test instance. For this reason, the test instance, specifically the medical data of the test instance, may be used as input "I," and a base learner model may be used to determine the model-specific indicator as output "O."

[0057] FIG. 5 schematically illustrates an exemplary method for generating multiple models according to the present disclosure. The method may be performed by device 100 or a different device. The method of FIG. 5 may be based on the method of FIG. 3. In particular, again, a base learner model may be determined using training data and a learner configuration file. As shown in FIG. 5, multiple different base learner models BLS1-BLSk may be determined. The base learner models BLS1-BLSk may differ from each other by at least one of the training data and the learner configuration file used to generate the respective base learner models. Each of the multiple base learner models BLS1-BLSk may be tested using a predefined (e.g., labeled) medical dataset to obtain characteristics of the respective models. The characteristics of the respective models may be, for example, the empirical performance of the base learner model (e.g., temporal performance, probability of false negative or false positive detection, etc.). The model characteristics may include characteristics associated with the training data used to generate the respective base learner models. The model characteristics may include the explainability of the respective base learner models. The multiple base learner models BLS1 to BLSk shown in FIG. 5 can be used as multiple models in the method of FIG.

[0058] 6 schematically illustrates an exemplary method for determining an indicator of a medical condition according to the present disclosure. The method may correspond to the method of FIG. 2. The method may be performed by the apparatus 100. A plurality of base learner models BLS1 to BLSk may be determined according to the method of FIG. 5.

[0059] The test instance, specifically the medical data of the test instance, can be used as input for a selector that can select at least one model from a plurality of base learner models BLS1 through BLSk. The selector can correspond to the selector described above in connection with the method of the first aspect. The selector can be defined by a selector configuration ("config") file. The selected model can then be used to determine the respective model-specific indicators. In the illustrated example, the base learner model BLS2 and the base learner model BLSk can be selected from a plurality of base learner models BLS1 through BLSk. In this example, as shown in FIG. 6, only the selected base learner models BLS2 and BLSk can be used to determine the respective model-specific indicators. The selection of the at least one model can be performed before the model-specific indicators are determined and before the indicators are determined.

[0060] Based on the model-specific indicators, a (e.g., final) indicator of the medical condition can be determined, for example, by or using an aggregator. The aggregator may correspond to the aggregator described above in connection with the method of the first aspect. The aggregator can determine or provide the (e.g., final) indicator based on the determined model-specific indicators, model characteristics, test instance metadata, and features extracted from the test instances. The aggregator can be defined by an aggregator configuration (“config”) file. It should be noted that the aggregator need not necessarily be a hardware component but may be embodied in software. The aggregator configuration file can define how inputs to the aggregator are combined into an output “O,” i.e., an indicator of the medical condition. The aggregator can include or be configured with an aggregation model as described herein above. The aggregation model can be defined by the aggregator configuration file. It should be noted that not all inputs shown in FIG. 6 can be used with the model-specific indicators to determine an indicator of the medical condition. For example, the index can be determined only based on the model-specific index and the test instance metadata.

[0061] As mentioned in connection with FIG. 2 , preferably, at least one of the base learner models BLS1 to BLSk may be selected based on at least one characteristic associated with the medical data of the test instance. The at least one characteristic associated with the medical data may include a medical image characteristic included in the medical data, corresponding to the “extracted features” shown in FIG. 6 , such as the slice thickness of the medical image, the contrast of the medical image, the imaging modality of the medical image (e.g., CT or MR), or a radiomics characteristic of the medical image. The radiomics characteristic may be extracted from the medical image using, for example, the software package “pyradiomics” available at https: / / github.com / Radiomics / pyradiomics. The at least one characteristic associated with the medical data may include one characteristic of the patient to which the medical data (and, for example, the test instance) pertains, corresponding to the “test instance metadata” shown in FIG. 6 . The test instance metadata may be included in the header of the medical image of the test instance. The characteristic may be, for example, the age, ethnicity, or gender of the patient to which the medical data pertains.

[0062] The selection step includes comparing, for each of the plurality of models, at least one feature associated with the medical data of the test instance with at least one feature associated with the training data used to generate the individual model. The comparison results in a suitability rating for each of the plurality of models. For example, a model trained on training images with a low slice thickness may have a lower suitability rating compared to test instance medical images with a higher slice thickness. A model trained on training images with a constant contrast may also have a lower suitability rating compared to test instance medical images with contrasts that deviate from the constant contrast by more than a predefined threshold. Similarly, a base learner model trained on training data of patients over 50 years old may have a lower suitability rating for test instance medical data related to 5-year-old patients. As a further example, a base learner model trained only on training data of Asian patients may have a lower suitability rating for test instance medical data related to patients of another ethnicity, such as African or European descent. Similarly, a base learner model trained only on training data of female patients may have a lower suitability rating for test instance medical data related to male patients.

[0063] The method may further include determining that a reliable determination of the indicators is not possible if at least one feature associated with the medical data of the test instance does not indicate the suitability of the at least one model, in which case a notification such as an audible or visual alert may be triggered to be output on the output device, for example by sending a corresponding trigger signal to the output device via the interface 106.

[0064] The indicator may be determined (e.g., by or using an aggregator) based on a majority vote of the model-specific indicators or based on an average aggregation for continuous numerical model-specific indicators. Other aggregation functions are possible. For example, if the indicators are highly sensitive, if one or more of the model-specific indicators indicate ill health, the indicator may be that the patient to whom the medical data of the test instance pertains is unhealthy.

[0065] The indicator may further be determined based on at least one of attributes selected from the results of the comparison step, the empirical performance of each model of the plurality of models (e.g., included in the model attributes), and the explainability of each model of the plurality of models (e.g., included in the model attributes). The explainability may be determined based on a learner configuration file or may be predetermined, and indicates how well a user can understand the functioning of the base learner (model). The explainability may be obtained from a database.

[0066] For example, only model-specific indicators having attributes that meet predetermined criteria may be taken into account (e.g., used or considered) for determining the (e.g., final) indicator. If any attributes do not meet the predetermined criteria, it may be determined that it is not possible to reliably determine an indicator, and a corresponding notification may be output as pointed out above.

[0067] Alternatively, a weighted average of the model-specific indicators can be determined as a (e.g., final) indicator of the medical condition. The weight of each model-specific indicator can be determined based on attributes, and in particular based on empirical performance. A weight w_i (e.g., summing to 1) can be defined for each learner model BLS1-BLSk, which can be directly proportional to the empirical performance of the base learner. For example, w_i can be directly proportional to exp(p_i), where p_i can be the empirical performance of base learner model i.

[0068] Further alternatively, the (e.g., final) indicator can be determined using an aggregation model, such as the aggregation model described above in this specification. In particular, the aggregator may include or be configured with the aggregation model described above in this specification and / or may be generated, for example, by a supervised learning algorithm. The supervised learning algorithm may be trained using the set of determined model-specific indicators, and optionally the set of model attributes, test instance metadata, and / or extracted features, and the generated aggregation model may be capable of providing an indicator of a medical condition based on the input of the "aggregator." The aggregation model training algorithm may be defined, for example, by an aggregator configuration file.

[0069] The indicator of a medical condition may be used to trigger the output of a notification on an output device informing a user about the indicator. For example, the interface 106 may output a trigger signal on a display, which then displays a visualization of areas of a medical image included in the medical data of the test instance that indicate the medical condition. Alternatively, a score or a binary visualization, such as red or green, may be displayed to inform the user of the indicator of the medical condition. In one example, the indicator itself may be visualized on the display. Alternatively, or additionally, a medical diagnosis may be determined based on the indicator of the medical condition. For example, if the indicator of the medical condition represents an abnormal volume in a medical image, further characteristics of the abnormal volume, such as image value, color value, volume size, etc., may be taken into account (e.g., used or considered) to determine a diagnosis, such as a diagnosis that the patient to whom the medical data relates has a brain tumor.

[0070] As will be apparent to one skilled in the art, other approaches to determining (e.g., final) indices based on model-specific indices may be possible. Below, examples of base learners and base learner models are described.

[0071] In general, supervised learning algorithms may require labeled training data. The labels may indicate whether a certain medical condition or set of medical conditions applies to the training data. The labeled training data can then be used to train the supervised learning algorithm so that it can predict the labels as model-specific indicators when applied to new data that has not been used for training. In one variation, the labeled training data (e.g., of at least one healthy patient and at least one unhealthy patient with a disease) can be used by the supervised learning algorithm to generate at least one of the base learner models BLS1 through BLSk.

[0072] Particularly in medical diagnosis tasks, the structure of the labels to be predicted can be hierarchical. As shown in Figure 7, there may be a higher-level label (also called a class label) for a certain class (e.g., "normal" corresponding to "healthy" or "abnormal" corresponding to "disease" or "unhealthy"). Some classes (e.g., the "disease" class) may have multiple subclasses (e.g., different subclasses D1-DZ for different diseases such as "lung cancer," "breast cancer," "prostate cancer," and "melanoma") that can be associated with corresponding lower-level labels (also called subclass labels).

[0073] When a model is generated by a supervised learning algorithm using labeled training data, the model may only be able to provide (e.g., determine) a model-specific indicator of whether a test instance is associated with one of the labels for which the model was trained. In this case, determining a model-specific indicator of a medical condition associated with the medical data of a test instance, where the medical condition is "healthy" or "disease," may be possible only if the training data represents a balanced amount of all possible diseases D1-DZ that can potentially be discovered. In other words, to train a supervised learning algorithm to enable correct prediction of a lower-level label of the subclass (the subclass "D2," which may correspond to the highlighted "lung cancer" in FIG. 7) or each higher-level class ("disease" in FIG. 7), labeled data for each possible subclass D1-DZ may need to be available. In particular, for diseases with very low prevalence (see amyotrophic lateral sclerosis (ALS) and nasopharyngeal carcinoma (NPC) in FIG. 8), such training data may be difficult to find or unavailable. As a result, it can be difficult in some cases to determine model-specific indicators of a medical condition using models generated by supervised learning algorithms.

[0074] A reliable machine-based indicator of medical condition "healthy" or "abnormal" (or "disease") may enable physicians to prioritize and appropriately time cases even before they are seen. As noted above, the methods described herein further include determining a medical diagnosis based on the indicator of the medical condition, thereby improving the physician's medical workflow.

[0075] In view of the above, a method for determining a model-specific indicator of a medical condition can be provided, as shown schematically in Figure 9. It should be noted that the model-specific indicator can be used directly as an indicator of the medical condition, and thus the method can be a method for determining an indicator of a medical condition. The method can be performed by the device 100.

[0076] The method includes step 902 of correlating portions of (e.g., the) medical image contained in (e.g., the) medical data of (e.g., the) test instance with portions of a reference image, for example using the first registration described above.

[0077] The method also includes step 904 of comparing image values ​​of at least one portion of the medical image with information associated with the correlated portion of the reference image to obtain an indication of the medical condition.

[0078] The information was generated, for example, by different devices, by: aligning a plurality of training images with a base image (e.g., using the second registration described above) and correlating a portion of each training image of the training images with a portion of the base image; determining an image value of at least one portion of each training image of the plurality of training images correlated with a portion of the base image, wherein the portion of the base image is assigned to the correlated portion of the reference image using a predetermined transformation (e.g., the third registration described above); and determining information based on the determined image value of the at least one portion of each training image of the plurality of training images.

[0079] The base image may be an atlas image generated based on multiple medical images. The base image may be determined by applying rigid registration of the multiple medical images to a common reference frame and averaging the image values ​​of all of these medical images. Applying rigid registration allows correlation of each portion of each patient image with the common reference frame. That is, the rigid registration may include a transformation matrix describing the transformation of the coordinate system of the patient image to a common coordinate system. Different rigid registrations may be used for different medical images among the multiple medical images. The base image may be an MR atlas image. CT images of the training data may be aligned to the MR atlas image using Mattes mutual information. In one variation, each of the training images may first be rigidly and affinely transformed to the base image before the SyN nonlinear transformation is applied. Each of the above registration steps may be repeated iteratively until convergence or a maximum number of iterations is reached. Similarly, each of these steps can be performed at different resolution levels, starting with a low resolution and then progressing to a higher resolution, to obtain a coarse registration that can then be refined into a finer registration.

[0080] A dataset (dimensions m × n × p) of multiple 3D CT images of a healthy patient can be mapped as training images to a base image, which can be a CT volume. This mapping can be performed by applying a registration between each of the multiple 3D CT images of the healthy patient (e.g., the second registration described above) to the base image. By applying the registration, each portion of each patient image can be correlated with a portion of the base image. That is, the registration can include a transformation matrix describing the transformation of the patient image coordinate system to the base image coordinate system and / or multiple transformations of different portions of the patient image to the base image coordinate system. Of course, different registrations can be used for different patient images. The registration can be determined using symmetric diffeomorphic image registration with cross-correlation and the SyN algorithm included in the Advanced Normalization Tools library, available, for example, at http: / / github.com / ANTsX / ANTs.

[0081] In a first implementation, the information may include or be a statistical distribution function of image values ​​of at least one portion of the plurality of training images. The information may be determined as follows: for each voxel location X, where 1≦a≦m, 1≦b≦n, and 1≦c≦p. a,b,c For , we can fit a statistical distribution function across all voxel values ​​of all training images. The distribution can be, for example, a Gaussian or a kernel density distribution (KDE). This allows us to fit an m × n × p distribution function p, estimated independently based on the training images. a,b,c can result.

[0082] The base learner model of the first implementation can use medical data to determine or provide a model-specific index of a medical condition as follows: That is, if a model-specific index of a medical condition is to be determined based on a medical image, the medical image can first be registered to a reference image that is registered to or is the base image using a predetermined transformation. If a voxel value of the medical image is higher than the upper q / 2 quartile or lower than the lower q / 2 quartile as defined by the corresponding distribution function, the voxel can be tagged or identified as abnormal. A percentile filter can be used to smooth the results, and thresholding can be used to obtain a binary segmentation mask. The binary segmentation mask can indicate portions of the test instance medical image that are abnormal. The binary segmentation mask can be determined as the model-specific index. An overall abnormality score for the test instance medical image can be determined as the model-specific index, for example, by counting the number of abnormal voxels in the medical image or dividing this number by the total number of voxels in the medical image.

[0083] In a second implementation, the information may include a mean image value of at least one portion of all of the plurality of training images and, optionally, a mean deviation of the image values ​​of at least one portion of all of the plurality of training images from the mean image value. The information may be determined as follows: for each voxel location X, where 1≦a≦m, 1≦b≦n, 1≦c≦p, a,b,c For each voxel location, the voxel values ​​of all training images can be obtained. For each voxel location, the average voxel value of all training images can be determined. In other words, the training images can be averaged voxel-wise. Prior to this, the training images can be normalized so that all voxel values ​​fall within a predetermined range (e.g., [0,1]). Furthermore, a mean error map can be calculated by calculating the average over all voxel-wise differences of all training images relative to the average voxel value.

[0084] The base learner model of the second implementation can use medical data to determine or provide a model-specific index of a medical condition as follows: If a model-specific index of a medical condition is to be determined based on a medical image, the medical image can first be registered to a reference image that is registered to or is the base image using a predetermined transformation. The absolute difference of the medical image relative to the average voxel value can then be determined for each voxel. A mean error map is subtracted from the absolute difference. A percentile filter can be used to smooth the results, and thresholding can be used to obtain a binary segmentation mask. The binary segmentation mask can indicate portions of the test instance medical image that are abnormal. The binary segmentation mask can be determined as the model-specific index. An overall abnormality score for the test instance medical image can be determined as the model-specific index, for example, by counting the number of abnormal voxels in the medical image or dividing this number by the total number of voxels in the medical image.

[0085] The method of Figure 9 can be combined with the methods described above in connection with Figures 2-6. In particular, at least one of the base learner models BLS1-BLSk described above can correspond to, for example, the density model described above, providing (e.g., having) the functionality of (e.g., having) the capability of) the method of Figure 9. The indicator of a medical condition determined by the method of Figure 9 can be used as a model-specific indicator in the methods described above in connection with Figures 2-6. In other words, at least one of the models included in the plurality of base learner models BLS1-BLSk can be generated by an unsupervised learning algorithm using unlabeled training data of healthy patients and can be configured to provide (e.g., determine) anomaly detection as a model-specific indicator of a medical condition.

[0086] As described above, the model-specific indicator of the medical condition and / or the (e.g., final) indicator of the medical condition may include at least one result selected from the probabilities of abnormalities for different portions of the medical images included in the medical data, and a numerical value describing the probability of abnormality of the overall medical data, where the numerical value is optionally derived from the probabilities of abnormalities for different portions of the medical images.

[0087] As should be apparent from the foregoing, the present disclosure provides techniques for determining indicators of a medical condition. At least one of the base learner models may be generated using unlabeled training data of healthy patients. Such base learner models may provide (e.g., determine) a model-specific indicator of a medical condition, which may be a score indicating whether a new, previously unseen test instance is in-distribution or out-of-distribution compared to the distribution of the training data. Thus, it is possible to generate at least one model using (only) readily available training data of healthy patients (for example).

[0088] Multiple base learner models may represent an ensemble. A (most suitable) subset of multiple base learner models can be selected based on the test instances at hand. In other words, the models to be included in the ensemble can be dynamically selected based on medical data. This may allow for a more reliable determination of model-specific indicators. The model-specific indicators of the selected base learner models can be aggregated to determine an indicator of the medical condition. Aggregation can further improve the reliability of the prediction (i.e., prediction of the determined indicators).

[0089] In particular, if the errors of the base learners are uncorrelated, combining them into one ensemble may enhance robustness. Furthermore, model-specific metrics can be combined in a way that explainsable base learners are preferred over unexplainable deep learning models. This may result in greater interpretability of the model-specific metrics. False negative predictions can be avoided by selecting at least one base learner model and determining metrics based on the model-specific metrics.

[0090] By using training data of a higher-level "normal" class (data of a class with no or few subclasses, e.g., training data of only healthy patients), a robust and reliable indication of whether the medical data of a test instance belongs to an in-distribution instance ("normal" or "healthy" instance) can be made possible. Through the proposed ensemble (rule-based or learning-based) coupling mechanism, specifically by determining the index based on model-specific indices, high predictive performance and increased prediction robustness can be guaranteed. The techniques described herein may be applicable to any type of diagnostic procedure and can provide predictions regarding higher-level classes without the need to use training data with detailed subclass labeling, or even without the need to use training data of out-of-distribution samples.

[0091] The advantages of the technology presented herein are believed to be fully understood from the foregoing description, and it will be apparent that various changes can be made in the form, construction, and arrangement of the exemplary embodiments thereof without departing from the scope of the disclosure or sacrificing all of its advantageous effects. Because the technology presented herein is susceptible to variation in many ways, it is recognized that the present disclosure should be limited only by the scope of the following claims.

Claims

1. 1. A method of processing medical data for determining an indicator of a medical condition, comprising: selecting (202) at least one model from a plurality of models (BLS1-BLSk) based on at least one characteristic associated with medical data of a test instance, each of the plurality of models (BLS1-BLSk) generated by a learning algorithm and configured to provide a model-specific indicator of a medical condition based on the medical data; determining (204) a model-specific indicator of each of the medical conditions based on the medical data using each of the at least one selected model; determining (206) the indicator of the medical condition based on the model-specific indicator; determining that a reliable determination of the indicator is not possible if the at least one characteristic associated with the medical data of the test instance does not indicate suitability of the at least one model; A method comprising:

2. The method of claim 1 , wherein the at least one characteristic associated with medical data includes a feature of a medical image contained in the medical data.

3. The method of claim 1 or 2, wherein the at least one characteristic associated with the medical data includes a characteristic of a patient to which the medical data relates.

4. 4. The method according to claim 1, wherein the selecting step comprises comparing, for each of the plurality of models (BLS1 to BLSk) individually, the at least one feature associated with the medical data of the test instance with at least one feature associated with training data used to generate the individual model.

5. 5. The method of claim 4, wherein the index is further determined based on at least one attribute selected from the results of the comparing step, the empirical performance of each of the plurality of models (BLS1 to BLSk), and the degree of explainability of each of the plurality of models.

6. 6. The method of any one of claims 1 to 5, wherein at least one of the models included in the plurality of models (BLS1 to BLSk) is generated by an unsupervised learning algorithm using unlabeled training data of healthy patients, and is optionally configured to provide anomaly detection as a model-specific indicator of a medical condition.

7. 7. The method according to any one of claims 1 to 6, wherein the model-specific indicator of the medical condition and / or the indicator of the medical condition comprises at least one result chosen from probabilities of abnormality for different parts of medical images contained in the medical data and a numerical value describing the probability of abnormality of the entire medical data, said numerical value optionally being derived from the probabilities of abnormality for said different parts of the medical image.

8. A method according to claim 4 or 5, wherein determining that a reliable determination of the index is not possible is based on the results of the comparison.

9. The method of claim 8, further comprising: triggering output of a notification on an output device; The method according to any one of claims 1 to 8, wherein the notification informs the user that a reliable determination of the indicator is not possible.

10. at least one of the models included in the plurality of models (BLS1 to BLSk) correlates a portion of a medical image included in the medical data with a portion of a reference image, and compares image values ​​of at least one portion of the medical image with information associated with the correlated portion of the reference image to obtain the model-specific indicator of the medical condition; Said information: aligning a plurality of training images with a base image to correlate portions of each of the training images with portions of the base image; determining an image value of at least one portion of each of the plurality of training images correlated with a portion of the base image, the portion of the base image being assigned to the correlated portion of the reference image using a predetermined transformation; determining the information based on the determined image values ​​of the at least one portion of each training image of the plurality of training images; The method of any one of claims 1 to 9, wherein the product is produced by

11. The method of claim 10 , wherein the information includes or is a statistical distribution function of image values ​​of the at least one portion of the plurality of training images.

12. 12. The method of claim 10 or 11, wherein the information comprises a mean image value of the at least one portion of all of the plurality of training images and, optionally, a mean deviation of the image values ​​of the at least one portion of all of the plurality of training images from the mean image value.

13. In an apparatus (100) including at least one processor (102) and at least one memory (104), the at least one memory (104) is a unit of the apparatus (100).

13. An apparatus storing instructions executable by said at least one processor (102) such that said at least one processor (102) is operable to perform the method of any one of claims 1 to 12.

14. A computer program product comprising program code portions for performing the method of any one of claims 1 to 12 when the computer program product is executed on one or more processors.

15. 15. The computer program product of claim 14 stored on one or more computer-readable recording media.

Images