Method and device for determining an interpretation diagram
Through deep neural network generation and interpretation of images and optimization of cost functions, the problem of insufficient information in medical diagnosis of retinal classification equipment is solved, and pixel-accurate image representation is achieved, which improves the accuracy of diagnosis.
Patent Information
- Application Number
- CN201980051444.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-08-03
- Filing Date
- 2019-07-03
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2039-07-03
AI Technical Summary
Existing retinal classification equipment is difficult to accurately display information on the decisive areas of the image in medical diagnosis, resulting in insufficient diagnostic results.
Through deep neural networks, interpreted images are generated to ensure accurate representation of pixels of the image, and selectively change the decisive pixels by optimizing the cost function to ensure that interpreted images do not change the classification results of the original image when classified, and edge and color information are preserved.
The pixel-accurate interpreted image representation is achieved, allowing human doctors to clearly identify the decisive areas in the image and improve the accuracy of medical diagnosis.
Smart Images

Figure CN112513878B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for determining an interpretation diagram, a computer program, one or more machine-readable storage media and one or more computers, a machine classification system and a display system. Background Art
[0002] The non-prepublished DE 10 2018 205 561 discloses a device for classifying retina images, the device comprising an artificial neural network having at least one convolutional layer,
[0003] wherein the convolutional layer comprises a plurality of trainable convolutional kernels, and an activation map is provided for each convolutional kernel, the activation map assigning output values to discrete positions in the input of the at least one convolutional layer, respectively, the output value being a measure of the local consistency of the input with the respective convolutional kernel,
[0004] wherein at least one classifier layer is connected after the at least one convolutional layer, said classifier layer mapping its input by means of trainable weights into a set of probabilities that the input belongs to one of a plurality of predefined classes (in particular whether the tissue is healthy or pathologically altered),
[0005] An evaluation unit is provided which is designed to determine a correlation map of the output of the at least one convolutional layer by weighted summation of an activation map obtained by the at least one convolutional layer and weights stored in the aggregation layer, wherein the correlation map displays areas of the tissue that are decisive for the evaluation.
[0006] Advantages of the present invention
[0007] In comparison, the method with the features of the invention has the advantage that it allows an improved determination of an interpretation map, which in particular allows a pixel-accurate representation, so that the information contained in the interpretation map is perceptible to humans.
[0008] Advantageous developments are the subject of the following description. Summary of the invention
[0009] In particular in systems for assisting medical diagnosis, in which possible diagnoses are identified with the aid of deep neural networks based on certain features of the image, it is important to clearly show the attending physician which parts of the image led to the diagnosis.
[0010] In a first aspect, the invention therefore relates to a method for determining an interpretation map of an image, i.e. an image region that is decisive for a classification result, wherein the image is classified (i.e. assigned to one of a plurality of predefined classes) with the aid of a deep neural network, in which interpretation map pixels of the image are changed depending on how much they contribute to the classification of the image, wherein the interpretation map is selected such that as small a subset of pixels of the image as possible is changed, this being done in such a way that when the interpretation map is fed to the deep neural network for classification it does not lead to the same classification result as the image as far as possible, wherein the interpretation map is selected such that in a feature map of the deep neural network the activations caused by the interpretation map then (i.e. when fed accordingly to the deep neural network) do not substantially exceed the activations caused by the image.
[0011] The fact that the interpretation image does not lead to the same classification result as the image as far as possible can mean, for example, that a probability is respectively assigned to each of the possible classification results and that the probability of identical classification results is minimized.
[0012] The change of pixels can be carried out, for example, in the following way: all those pixels of the image that are decisive for the classification of the image are changed. These pixels can be highlighted or deleted, for example.
[0013] This means that provision is made for both the image and also the interpretation map to be fed to the deep neural network, wherein the classifications obtained in each case as output variables of the deep neural network should be as different as possible.
[0014] Here, a deep neural network is generally a neural network including a plurality of layers.
[0015] A particularly accurate representation results since the explanation map is created at the size of the input image and not at the size of the activation map of the (especially the last) convolutional layer or the size of the output variable, as the explanation map would have to be scaled up to the size of the input image in the described alternative, while retaining relevant edge and color information.
[0016] In one embodiment, it can be provided that the explanation graph is selected such that the cost function L (English: loss function) is optimized. In other words, the proposed method is based on an optimization problem with auxiliary conditions, i.e., for example
[0017]
[0018] Equation (1) can advantageously be solved with the aid of a mask b, which can be multiplicative, for example, sx = x·b, where b∈[0,1] is optimized, advantageously using the starting condition b(0)=1. Alternatively, the mask b can also be additive, i.e., s x =s+b, where b is optimized with the starting condition b(0)=0. If b* is used to denote the optimized mask, the solution to the optimization problem from equation (1) is or
[0019] In the optimization problem, we start from the image x and find an explanation graph or representation that satisfies certain properties. These properties can be represented particularly simply and flexibly using the cost function.
[0020] So the image x and representation The relationship between x Initial state (or by the corresponding starting conditions of mask b) and by the auxiliary conditions of the optimization. A possible implementation is to represent contains only the features that are also present in image x. Therefore, the initial state can be Advantageously, an auxiliary condition that should be satisfied for each ReLU layer of the deep neural network may be h l (s x )≤h l (x), where h l is the activation of the first ReLU layer of the deep neural network (of course, other activation functions can also be used instead of the ReLU function).
[0021] The cost function advantageously has multiple (additive) components, namely L = L h +L w +L c Of course, these additive components can be weighted differently by parameters, but these parameters can also be integrated into the definition of these components. When solving the optimization problem from equation (1), these weighted parameters can optionally be adapted together. The first component L can be selected h = l h (s x ), so that it represents how many pixels of the image have been changed. For example, this component can be expressed by the L1 norm l h (s x )=||s x Alternatively, this component can be used for s by multiplying the L1 norm of the mask b by ||1-b||1 x = x·b(b∈[0,1]). This component then makes all irrelevant pixels from x remain unchanged in s.x middle.
[0022] Alternatively or additionally, the cost function may have a second component which characterizes how likely it is that the classification result obtained for the explanation image corresponds to the classification result obtained for the image. This can be achieved by means of the component L w = l w (f(x|θ),f(s x |θ)), where f describes the deep neural network or a part of the deep neural network. The parameters of the deep neural network are represented by θ. The output of the deep neural network, i.e., the classification, is obtained by f(x|θ) or f(s x |θ) indicates.
[0023] Alternatively, the second component can be completed region by region, in particular pixel by pixel, by a function that characterizes how much the respective pixel contributes to the classification result obtained for the image. For example, the logit function of category i can be used This function is given by That is, the second component is given by the probability of class i (in the case of minimizing the probability of class i). Alternatively, the probability can be given by a function that determines all pixels that activate feature i. If represents the i-th feature map of the l-th layer, then the associated term of the cost function can be obtained by Other paradigms are possible, such as the L2 paradigm.
[0024] The auxiliary condition that the activation by the interpretation image does not substantially exceed the activation by the image can be achieved particularly easily by means of the optimization approach via the third component L in the cost function. c The third component penalizes such an excess, i.e., for example, L c =||maximum(h l (s x )-h l (x),0)||1. Other normalizations are possible, such as the L2 normalization.
[0025] Alternatively, the auxiliary condition can be satisfied by: when determining the interpretation map, the activation h in the characteristic map is set to l (s x ) is set to a predeterminable value, in particular to a maximum permissible value according to the auxiliary condition, for example by h′ l (s x )=min(h l (x), h l (s x)). If a gradient-based optimization method is used, then δ′ l =δ l ·[h l (s x )<h l (x)] gives the corresponding gradient δ in back propagation l The symbol [...] here usually means if the condition is met, then value = 1, if the condition is not met, then value = 0.
[0026] Alternatively, the auxiliary condition can be satisfied in the following manner: when using a gradient-based optimization method, if the explanation graph s is used in a layer with an activation function (particularly a ReLU layer) of the deep neural network x Activation h l (s x ) exceeds the activation h of the image x l (x), i.e., for example, δ′ l =δ l ··[h l (s x )<h l (x)], then the feature map h of this layer l Back-propagated gradient δ l (ie the gradient present at the output of the activation function) is set equal to a predefinable gradient value, in particular equal to zero.
[0027] The interpretation image thus determined can now advantageously be determined and displayed, in particular next to the image. Thus, by means of a positionally realistic resolution and the selection of only irrelevant pixels, a representation can be generated that can be examined by a human expert, in particular a doctor. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] The embodiments of the present invention are explained in more detail below with reference to the accompanying drawings. In the accompanying drawings:
[0029] Figure 1 The structure of an embodiment of the present invention is schematically shown;
[0030] Figure 2 The flow chart shows the process of a method according to one aspect of the present invention. DETAILED DESCRIPTION
[0031] In a preferred embodiment of the present invention, Figure 1A machine classification system 40 is shown, which is arranged to carry out the invention. The machine classification system 40 comprises a computer 45 and a machine-readable storage medium 46. The machine classification system further comprises a machine learning system 60, which determines an output signal y from an image x fed to it. The image x may be determined, for example, from a sensor signal S determined by an imaging sensor 30 via an optional input interface 50.
[0032] The machine learning system determines the classification f(x|θ) of the image x with the aid of a deep neural network f. The deep neural network f is parameterized using parameters θ, which are stored in a parameter memory Q. The machine learning system 60 also uses, for example, Figure 2 The method shown in Explanation diagram Together with the classification f(x|θ), an output signal y is formed, from which the control signal A is determined together with the image x, for example using the optional output interface 80. The control signal A may include, for example, the interpretation image The classification f(x|θ) and the image x are then transmitted to the display unit 10, which includes, for example, two monitors 11 and 12. The image x is transmitted to the first monitor 11 and displayed, and the interpretation image The data are sent to the second monitor 12 and displayed. Optionally, the classification f(x|θ) may also be displayed.
[0033] Figure 2 shows the interpretation diagram used to determine Implementation of the method. First (1000) a deep neural network f is trained, for example to be able to distinguish between different clinical images in an image. Then (1100) a classification f(x|θ) of the image x is determined for the image x fed to the machine learning system 60 with the aid of the deep neural network f. These steps are not necessarily part of the method according to the invention, but rather determine the context of the method and can, for example, be performed in advance by a separate unit.
[0034] Then (1200) the initial value of the explanation map is set For example, Then (1300) solves the optimization problem defined by equation (1), preferably using a gradient descent method, to determine the explanation graph
[0035] Then (1400), the explanation diagram The image x is supplied to the monitors 11, 12 and displayed as described on the monitors 11, 12. The method is thus terminated.
[0036] In another embodiment, the imaging sensor 30 is implemented in the manufacturing facility.The method can then be used to detectably display reject portions to a human operator via the display unit, thereby enabling the operator to identify via the display on the display unit 10 whether reject portions are actually present.
[0037] In another embodiment, an arbitrary image acquisition unit is provided instead of the imaging sensor 30, which may be the imaging sensor 30 or, for example, load and provide images from a database. The method can then be used to check the operability of the deep neural network f, so that a human developer can analyze via a display on the display unit 10 which image areas lead to possible misclassifications.
[0038] In yet another embodiment, the imaging sensor 30 is part of a surveillance system. By displaying on the display unit 10, a human expert is enabled to identify whether an abnormal scene (e.g., an abandoned suitcase) identified by the machine learning system 60 actually exists.
[0039] It is understandable to those skilled in the art that the computer 46 may be provided by multiple computers, and the machine-readable storage medium 45 may be provided by multiple machine-readable storage media.
Claims
1. An explanation graph for determining image x in which all pixels of the image x that are decisive for the classification of the image x determined by the deep neural network f are changed in the interpretation image, The explanation diagram Choose , so that the optimization problem with auxiliary conditions for the cost function L is optimized: in: The optimization problem is solved with the aid of a mask b, which is either multiplicative, i.e. s x = x·b, where the optimization b∈[0,1] is either additive, i.e., s x =x+b, where b is optimized with the initial condition b(0)=0; The explanation diagram is chosen so that the subset of pixels that are changed in the image x is minimized. If the explanation image is transmitted to the deep neural network f for classification, then the explanation graph does not lead to the same classification result as the image x The method is to assign a probability p to each possible classification result and minimize the probability of the same classification result, and make the explanation graph Activation caused by will not exceed the activation h caused by the image x l (x), The cost function L has a first component L h , where the first component L h Given by the L1 paradigm: h (s x )=||s x -x||1, the first component represents how many pixels of the image x have been changed, Wherein the cost function L has a second component L w , wherein the second component is pixel by pixel by characterizing the interpretation map The classification result corresponds to the probability function of the classification result of the image x, and The auxiliary condition is h l (s x )≤h l (x), that is, for all activation function layers, through the explanation graph Activation of the lth activation function layer of the deep neural network f No more than the activation h of the lth activation function layer of the deep neural network f by the image x l (x), The auxiliary condition is implemented by implementing a third component Lc in the cost function L, which penalizes such an excess. Or it can be realized by the following method: when determining the explanation diagram When the activation function of the corresponding lth activation function layer is Set to a pre-set value, Or it can be achieved by using a gradient-based optimization method and if in layer l of the deep neural network the explanation graph Activation Over the activation h of the image x l (x), then the activation of this layer The back-propagated gradient δ l Set to a predeterminable gradient.
2. The method according to claim 1, wherein: The predeterminable value is the maximum permissible value h of the auxiliary condition. l (x).
3. The method according to claim 1, wherein: The predefinable gradient is equal to zero.
4. A method for displaying an image, wherein: Determining an interpretation diagram using the method according to any one of claims 1 to 3 And display the explanation diagram 5. The method according to claim 4, wherein: In the explanation diagram The image x is also displayed alongside.
6. A computer program product arranged to perform the method according to any one of claims 1 to 5.
7. One or more machine-readable storage media (45) having stored thereon a computer program product according to claim 6.
8. A machine sorting system (40) arranged to perform the method according to any one of claims 1 to 3.
9. A display system (13) arranged to perform the method according to any one of claims 4 or 5.
