Method and apparatus for processing of multi-modal data
A method for training a machine learning model that is able to establish links between data of different modalities by creating a joint representation. In particular, application of the method to medical data including electronic medical records and medical images and/or other medical data. The trained machine learning model can among others fulfil tasks such as autocompletion of incomplete data, detection of uncertain and/or spurious data, generation of probable data and other tasks.
1 . A method comprising:
obtaining input text data indicative of at least partially edited source text data for training of a machine learning model and input image data indicative of at least partially edited source image data for training of the machine learning model;
obtaining test text data indicative of at least partially unedited source text data for training of the machine learning model and test image data indicative of at least partially unedited source image data for training of the machine learning model;
generating a joint representation of the input text data and the input image data from the input text data and the input image data depending at least on one parameter influencing the machine learning model, wherein the joint representation enables deriving output text data and output image data;
generating output text data indicative of at least a part of the unedited source text data and output image data indicative of at least a part of the unedited source image data based, at least in part, on the joint representation of the input text data and the input image data, and depending at least on one parameter influencing the machine learning model;
evaluating the performance of the machine learning model by computing a first loss function indicative of a first deviation between the output text data and the test text data,
evaluating the machine learning model by computing a second loss function indicative of a second deviation between the output image data and the test image data, and
changing at least one parameter of the machine learning model to influence the output text data and/or the output image data generated by the machine learning model in order to reduce the first deviation between the output of the first loss function and a desired output of the first loss function and/or the second deviation between the output of the second loss function and a desired output of the second loss function.
2 . The method of claim 1 , wherein:
the at least partly edited source image data comprises at least partially cropped and/or at least partially corrupted source image data wherein corrupting is one or more of masking, rearranging, distorting, rotating, mirroring, coloring, changing contrast, changing brightness, pixelating, smoothing, and adding noise and/or the at least partly edited source text data comprises at least partially cropped and/or at least partially corrupted source text data wherein corrupting is one or more of masking, rearranging text parts, deleting text parts, replacing text parts with new text, wherein text parts are one or more of letters, numbers, words, sentences, paragraphs or pages.
3 . The method of claim 1 , wherein:
the test text data comprises at least partially cropped source data and/or the test image data comprises at least partially cropped source image data.
4 . The method of claim 1 , wherein:
the changing of the at least one parameter of the machine learning model is dependent on the magnitude of the first deviation crossing a first threshold and/or on the magnitude of the second deviation crossing a second threshold.
5 . The method of claim 4 , wherein:
the method is iteratively repeated until a number of iterations or a threshold-crossing of at least one of the loss function outputs is met wherein at least one of the input text data and the input image data is at least partially altered between repetitions.
6 . The method of claim 1 , wherein:
the machine learning model comprises an artificial neural network, wherein changing of the at least one parameter of the machine learning model is achieved by backpropagation.
7 . The method of claim 1 , wherein:
the input text data and/or the input image data comprise information about a patient,
the input text data at least partially represents an electronic medical record of the patient and/or comprises information about vital parameters, physiological measurements, prescribed, recommended and/or other medications, and/or the input image data represents at least one image of at least one body region of the patient, wherein the image was acquired by photography, X-ray radiography, computerized tomography, fluoroscopy, magnetic resonance imaging, functional magnetic resonance imaging, ultrasonography, endoscopy, elastography, tactile imaging, thermography, microscopy, positron emission tomography and/or any other imaging technique.
8 . The method of claim 1 , wherein generating a joint representation of both the input text data and the input image data comprises:
encoding the input text data using a first encoding branch to an encoded representation of the input text data;
encoding the input image data using a second encoding branch to an encoded representation of the input image data; and
combining the encoded representation of the input text data and the encoded representation of the input image data to a joint representation of the input text data and the input image data, wherein both, the representation of input text data and the representation of input image data, are vectors, matrices and/or tensors, wherein combining encoded representations preferably comprises one or more of concatenating encoded representations, multiplying encoded representations, adding encoded representations, convoluting encoded representations.
9 . The method of claim 1 , wherein generating output text data and output image data from the joint representation of the input text data and the input image data comprises:
generating the output text data from the joint representation of both the input text data and the input image data using a first decoding branch; and
generating the output image data from the joint representation of both the input text data and the input image data using a second decoding branch.
10 . The method of claim 1 , further comprising:
generating a first representation based on first input text data and/or first input image data, the first representation being a representation of the input text data, a representation of the input image data or a joint representation;
generating a second representation based on second input text data and/or second input image data, the second representation being a representation of the input text data, a representation of the input image data or a joint representation; and
comparing the first representation and the second representation using a representation deviation metric.
11 . The method of claim 1 , further comprising:
changing at least one parameter of the machine learning model to influence the representation deviation metric in a manner at least partially depending on whether the first input text data and the second input text data stem from the same source text data and/or whether the first input image data and the second input image data stem from the same source image data.
12 . The method of claim 1 , further comprising:
providing new input text data and/or new input image data to the machine learning model,
wherein the new input text data comprises information about a patient, an electronic medical record of a patient or multiple patients, information about vital parameters, physiological measurements, prescribed, recommended and/or other medications, and/or
wherein the new image data represents at least one image of at least one body region of the patient, wherein the image was acquired by photography, X-ray radiography, computerized tomography, fluoroscopy, magnetic resonance imaging, functional magnetic resonance imaging, ultrasonography, endoscopy, elastography, tactile imaging, thermography, microscopy, positron emission tomography and/or any other imaging technique,
generating new output text data and/or new output image data using the trained machine learning and/or the apparatus.
13 . The method of claim 12 , further comprising:
repeating at least the steps of providing the new input text data and/or the new input image data and generating new output text data and/or new output image data at least once while randomly varying the new input text data and/or new input image data,
comparing the new output text data of different repetitions and/or comparing the new output image data of different repetitions,
quantifying the variability between repetitions of parts of the new output image data and/or quantify the variability between repetitions of parts of the new output text data,
identify parts of the new output image data and/or parts of the new output text data that are uncertain based on the quantifying of their variability.
14 . An apparatus comprising at least one processor and at least one memory including computer program code, the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to:
obtain at least one of input text data and/or input image data, and
generate output text data and/or output image data at least partially using a trained machine learning model, wherein training the machine learning model comprises:
obtaining input text data indicative of at least partially edited source text data for training of the machine learning model and input image data indicative of at least partially edited source image data for training of the machine learning model;
obtaining test text data indicative of at least partially unedited source text data for training of the machine learning model and test image data indicative of at least partially unedited source image data for training of the machine learning model;
generating a joint representation of the input text data and the input image data from the input text data and the input image data depending at least on one parameter influencing the machine learning model, wherein the joint representation enables deriving output text data and output image data;
generating output text data indicative of at least a part of the unedited source text data and output image data indicative of at least a part of the unedited source image data based, at least in part, on the joint representation of the input text data and the input image data, and depending at least on one parameter influencing the machine learning model;
evaluating the performance of the machine learning model by computing a first loss function indicative of a first deviation between the output text data and the test text data,
evaluating the machine learning model by computing a second loss function indicative of a second deviation between the output image data and the test image data, and
changing at least one parameter of the machine learning model to influence the output text data and/or the output image data generated by the machine learning model in order to reduce the first deviation between the output of the first loss function and a desired output of the first loss function and/or the second deviation between the output of the second loss function and a desired output of the second loss function,
wherein:
the output image data is at least partially dependent on the input text data, and/or
the output text data is at least partially dependent on the input image data.