Deep learning-based skin cancer detection method

Abstract

The invention provides a skin cancer detection method based on deep learning. The method comprises the following steps: collecting signal data of different skin parts of a patient; designing a signal preprocessing module to preprocess the signal data; sending the preprocessed signal data into a designed detection network module; designing a loss function to obtain a skin cancer detection model; and performing skin cancer detection on the skin of the patient by using the trained skin cancer detection model. The features of different skin parts of the patient are extracted by using a deep learning technology and classified and analyzed, automatic detection and diagnosis of skin cancer are realized, rapid and accurate cancer detection can be performed on the skin of the patient, a doctor is helped to perform early diagnosis and treatment, and the treatment effect and the survival rate of the patient are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of skin cancer detection, and more specifically, to a skin cancer detection method based on deep learning. Background Art

[0002] Skin cancer is a common malignant tumor, and early detection and diagnosis are crucial for the treatment and prognosis of skin cancer. Traditional skin cancer detection methods usually rely on naked-eye observation and pathological examination, which may require obtaining skin tissue samples and have the disadvantages of invasiveness and long time consumption. Therefore, it is very important to find a non-invasive, fast, convenient and rapid skin cancer detection method.

[0003] The skin cancer detection method based on deep learning uses impedance technology to obtain measurement data of skin resistance, and takes it as input to be trained and predicted through a deep learning model. The deep learning model can learn and extract features from a large amount of data and has strong pattern recognition and classification capabilities. By processing and analyzing the resistance measurement data, the deep learning model can learn the differential features between skin cancer and normal skin and perform accurate classification and diagnosis.

[0004] In the prior art, Chinese Patent No. CN202010097921 discloses "a skin cancer disease detection method based on unsupervised learning". This detection method trains through an autoencoder network by constructing a new neural network containing a clustering layer. However, this method is dependent on data quality, has difficulty in selecting hyperparameters, requires a long training time and high computing resources, and cannot perform fast and timely detection. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art or related technologies.

[0006] To this end, the object of the present invention is to provide a skin cancer detection method based on deep learning.

[0007] To achieve the above object, the technical solution of the present invention provides a skin cancer detection method based on deep learning. The skin cancer detection method based on deep learning includes: Step S1: Collect signal data of different skin parts of a patient; Step S2: Design a signal preprocessing module to preprocess the signal data; Step S3: Send the preprocessed signal data into a designed detection network module; Step S4: Design a loss function to obtain a skin cancer detection model; Step S5: Use the trained skin cancer detection model to detect the skin of the patient for skin cancer.

[0008] Preferably, the signal data includes: impedance data and phase angle data of different skin parts of the patient; and the patients include patients with normal skin, patients with benign skin, and patients with malignant skin.

[0009] Preferably, the signal preprocessing module includes a signal double-layer denoising processing module and a signal feature enhancement module; and step S2 specifically includes: Step S2.1: Through the signal double-layer denoising processing module, perform double-layer denoising processing on two columns of data in the original signal matrix X corresponding to the signal data; wherein, the two columns of data are respectively: impedance data and phase angle data; Step S2.2: Send the matrix X after double-layer denoising processing m to the signal feature enhancement module to obtain the matrix X e ; Step S2.3: Concatenate the original signal matrix X and the matrix X e to obtain the matrix X D .

[0010] Preferably, step S2.1 specifically includes: Step S2.11: Perform the first-layer denoising processing on two columns of data in the original signal matrix X corresponding to the signal data; the calculation formula corresponding to the first-layer denoising processing is:

[0011]

[0012] In formula (1), X sij represents the first-layer denoising processing value of matrix X m ; X ij represents the data of the i-th row and j-th column of X; |min(X j )| represents the absolute value of the minimum value of the j-th column; |max(X j )| represents the absolute value of the maximum value of the j-th column;

[0013] Step S2.12: Perform the second-layer denoising processing on X sij ; the calculation formula corresponding to the second-layer denoising processing is:

[0014]

[0015] In formula (2), X mij represents the second-layer denoising processing value of matrix X m ; max represents taking the larger value of the two values;

[0016] Step S2.2 specifically includes: Send X mij to the signal feature enhancement module to obtain the matrix X e ; the calculation formula corresponding to the signal feature enhancement module is:

[0017] Xeij = (softmax(w × relu(X mij + b))) × X mij + X mij (3)

[0018] In formula (3), X eij represents the value of matrix X e ; softmax represents the softmax activation function; relu represents the relu activation function; w and b represent learnable parameters; and

[0019] In step S2.3, the calculation formula for splicing the original signal matrix X and the matrix X e is as follows:

[0020] X D = reshape(X, X e )(4)

[0021] In formula (4), reshape represents the horizontal splicing operation.

[0022] Preferably, in step S3, the detection network module includes: a first depth feature extraction layer, a second depth feature extraction layer, a first temporal feature extraction layer, a second temporal feature extraction layer, a first adaptive spatio-temporal feature fusion layer, a second adaptive spatio-temporal feature fusion layer, a global enhancement layer, and a local enhancement layer; and step S3 specifically includes: Step S3.1: Send the matrix X D of the signal preprocessing module into the first depth feature extraction layer and the first temporal feature extraction layer respectively, and then fuse the feature matrices output by the first depth feature extraction layer and the first temporal feature extraction layer to obtain the matrix X D1 ; Step S3.2: Send the matrix X D1 into the first adaptive spatio-temporal feature fusion layer to obtain the matrix X TS1 ; Step S3.3: Send the matrix X TS1 into the second depth feature extraction layer and the second temporal feature extraction layer respectively, and then fuse the feature matrices output by the second depth feature extraction layer and the second temporal feature extraction layer to obtain the matrix X D2 ; Step S3.4: Send the matrix X D2 into the second adaptive spatio-temporal feature fusion layer to obtain the matrix X TS2 ; Step S3.5: Send the matrix X TS2They are respectively sent into the global enhancement layer and the local enhancement layer to obtain the feature matrices respectively output by the global enhancement layer and the local enhancement layer; Step S3.6: Fuse the feature matrices respectively output by the global enhancement layer and the local enhancement layer, and then sequentially pass the fused feature matrix through a fully connected layer and a softmax layer to obtain the corresponding skin cancer detection result.

[0023] Preferably, in step S3.1, the first depth feature extraction layer includes: a first 1×1 depthwise separable convolution, a first average pooling layer, a first sigmoid activation function, a second 1×1 depthwise separable convolution, a second average pooling layer, and a second sigmoid activation function connected in sequence; the first temporal feature extraction layer includes: a first bidirectional GRU recurrent neural network, a first 1×1 ordinary convolution, a first relu activation function, a second bidirectional GRU recurrent neural network, a second 1×1 ordinary convolution, and a second relu activation function connected in sequence;

[0024] In step S3.2, the first adaptive spatio-temporal feature fusion layer includes a first adaptive spatio-temporal attention mechanism module, a first normalization layer, and a third relu activation function; wherein, the calculation formula adopted by the first adaptive spatio-temporal attention mechanism module is as follows:

[0025]

[0026] In formula (5), C D1 represents the number of channels of matrix X D1 ; H D1 represents the height of matrix X D1 ; W D1 represents the width of matrix X D1 ; softmax represents the softmax activation function;

[0027] In step S3.3, the second depth feature extraction layer includes: a first 3×3 depthwise separable convolution, a first max pooling layer, a third sigmoid activation function, a second 3×3 depthwise separable convolution, a second max pooling layer, and a fourth sigmoid activation function connected in sequence; the second temporal feature extraction layer includes: a third bidirectional GRU recurrent neural network, a first 3×3 ordinary convolution, a fourth relu activation function, a fourth bidirectional GRU recurrent neural network, a second 3×3 ordinary convolution, and a fifth relu activation function connected in sequence;

[0028] In step S3.4, the second adaptive spatio-temporal feature fusion layer includes: a second adaptive spatio-temporal attention mechanism module, a second normalization layer, and a sixth relu activation function connected in sequence; wherein, the calculation formula adopted by the second adaptive spatio-temporal attention mechanism module is as follows:

[0029]

[0030] In formula (6), C D2 represents the number of channels of matrix X D2 ; H D2 represents the height of matrix X D2 ; W D2 represents the width of matrix X D2 ; softmax represents the softmax activation function;

[0031] In step S3.5, the global enhancement layer specifically includes: a 5×5 ordinary convolution connected in sequence, a 7×7 ordinary convolution, and a global average pooling layer; the local enhancement layer specifically includes: a 1×1 ordinary convolution connected in sequence, a third 3×3 ordinary convolution, and a third max pooling layer.

[0032] Preferably, the calculation formula of the loss function used in the detection network module is as follows:

[0033]

[0034] In formula (7), L represents the number of samples; y i represents the true label of the i-th sample; p i represents the probability of the predicted label of the i-th sample; ∈ is a positive smoothing term.

[0035] Preferably, step S5 specifically includes: inputting the signal data of different skin parts of a certain patient collected into the trained skin cancer detection model, and automatically obtaining the skin cancer detection result of the patient, so as to realize rapid and accurate cancer detection of the skin of the patient.

[0036] Advantages of the present invention:

[0037] (1) The skin cancer detection method based on deep learning provided by the present invention preprocesses the collected data by designing a signal preprocessing module, providing more reliable input data for subsequent skin cancer detection.

[0038] (2) The skin cancer detection method based on deep learning provided by the present invention can use deep learning technology to extract the features of different skin parts of the patient and perform classification analysis by sending the preprocessed data into the designed detection network module, realizing automatic detection and diagnosis of skin cancer.

[0039] (3) The skin cancer detection method based on deep learning provided by the present invention can use the trained detection model to perform rapid and accurate cancer detection on the skin of the patient, helping doctors with early diagnosis and treatment, and improving the treatment effect and survival rate of the patient.

[0040] Additional aspects and advantages of the present invention will become apparent in the following description or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic flowchart of a deep learning-based skin cancer detection method according to an embodiment of the present invention is shown;

[0042] Figure 2 A schematic diagram of a data preprocessing module of a deep learning-based skin cancer detection method according to an embodiment of the present invention is shown;

[0043] Figure 3 A schematic flowchart of a deep learning-based skin cancer detection method according to another embodiment of the present invention is shown. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] In order to be able to more clearly understand the above objects, features, and advantages of the present invention, as Figures 1 to 3 shown, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments may be combined with each other.

[0045] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the limitations of the specific embodiments disclosed below.

[0046] Figure 1 A schematic flowchart of a deep learning-based skin cancer detection method according to an embodiment of the present invention is shown. As Figure 1 shown, the deep learning-based skin cancer detection method includes: collecting signal data of different skin parts of a patient; designing a signal preprocessing module to preprocess the signal data; sending the preprocessed signal data into a designed detection network module; designing a loss function to obtain a skin cancer detection model; and using the trained skin cancer detection model to detect the skin of the patient for skin cancer to obtain a detection result of normal skin, a detection result of benign skin, or a detection result of malignant skin of the patient.

[0047] In this embodiment, the skin cancer detection method based on deep learning provided by the present invention preprocesses the collected data by designing a signal preprocessing module, providing more reliable input data for subsequent skin cancer detection. Further, by sending the preprocessed data into the designed detection network module, deep learning technology can be used to extract the features of different skin parts of the patient and perform classification analysis, realizing the automatic detection and diagnosis of skin cancer. Further, by using the trained detection model, rapid and accurate cancer detection of the patient's skin can be carried out, helping doctors with early diagnosis and treatment, and improving the treatment effect and survival rate of the patient.

[0048] In one embodiment of the present invention, the signal data includes: impedance data and phase angle data of different skin parts of the patient; and the patient includes patients with normal skin, patients with benign skin, and patients with malignant skin.

[0049] In one embodiment of the present invention, the signal preprocessing module includes a signal double-layer denoising processing module and a signal feature enhancement module; and as Figure 2 shown, step S2 specifically includes: Step S2.1: Through the signal double-layer denoising processing module, perform double-layer denoising processing on two columns of data in the original signal matrix X corresponding to the signal data; where the two columns of data are respectively: impedance data and phase angle data; Step S2.2: Send the matrix X after double-layer denoising processing m into the signal feature enhancement module to obtain matrix X e ; Step S2.3: Concatenate the original signal matrix X and the matrix X e to obtain matrix X D .

[0050] In one embodiment of the present invention, step S2.1 specifically includes: Step S2.11: Perform the first-layer denoising processing on two columns of data in the original signal matrix X corresponding to the signal data; the calculation formula for performing the first-layer denoising processing is:

[0051]

[0052] In formula (1), X sij represents the first-layer denoising processing value of matrix X m ; X ij represents the data of the i-th row and j-th column of X; |min(X j )| represents the absolute value of the minimum value of the j-th column; |max(X j )| represents the absolute value of the maximum value of the j-th column;

[0053] Step S2.12: X sijPerform the second - layer denoising process; the calculation formula corresponding to the second - layer denoising process is:

[0054]

[0055] In formula (2), X mij represents matrix X m the second - layer denoising value; max represents taking the larger value of two values; Step S2.2 specifically includes: sending X mij into the signal feature enhancement module to obtain matrix X e ; the calculation formula corresponding to the signal feature enhancement module is:

[0056] X eij =(softmax(w×relu(X mij +b)))×X mij +X mij (3)

[0057] In formula (3), X eij represents the value of matrix X e ; softmax represents the softmax activation function; relu represents the relu activation function; w and b represent learnable parameters; and

[0058] In step S2.3, the formula for concatenating the original signal matrix X and matrix X e is:

[0059] X D =reshape(X,X e )(4)

[0060] In formula (4), reshape represents the horizontal concatenation operation.

[0061] In an embodiment of the present invention, as Figure 3 shown, in step S3, the detection network module includes: a first depth feature extraction layer, a second depth feature extraction layer, a first temporal feature extraction layer, a second temporal feature extraction layer, a first adaptive spatio - temporal feature fusion layer, a second adaptive spatio - temporal feature fusion layer, a global enhancement layer, and a local enhancement layer; and step S3 specifically includes: Step S3.1: Send the matrix X D of the signal pre - processing module into the first depth feature extraction layer and the first temporal feature extraction layer respectively, and then fuse the feature matrices output by the first depth feature extraction layer and the first temporal feature extraction layer to obtain matrix X D1 ; Step S3.2: Send matrix X D1 into the first adaptive spatio - temporal feature fusion layer to obtain matrix X TS1; Step S3.3: Send the matrix X TS1 into the second depth feature extraction layer and the second temporal feature extraction layer respectively, and then fuse the feature matrices output by the second depth feature extraction layer and the second temporal feature extraction layer to obtain the matrix X D2 ; Step S3.4: Send the matrix X D2 into the second adaptive spatio-temporal feature fusion layer to obtain the matrix X TS2 ; Step S3.5: Send the matrix X TS2 into the global enhancement layer and the local enhancement layer respectively to obtain the feature matrices output by the global enhancement layer and the local enhancement layer respectively; Step S3.6: Fuse the feature matrices output by the global enhancement layer and the local enhancement layer respectively, and then sequentially pass the fused feature matrix through a fully connected layer and a softmax layer to obtain the corresponding skin cancer detection result.

[0062] In an embodiment of the present invention, as Figure 3 shown, in step S3.1, the first depth feature extraction layer includes: a first 1×1 depthwise separable convolution, a first average pooling layer, a first sigmoid activation function, a second 1×1 depthwise separable convolution, a second average pooling layer, and a second sigmoid activation function connected in sequence; the first temporal feature extraction layer includes: a first bidirectional GRU recurrent neural network, a first 1×1 ordinary convolution, a first relu activation function, a second bidirectional GRU recurrent neural network, a second 1×1 ordinary convolution, and a second relu activation function connected in sequence;

[0063] In step S3.2, the first adaptive spatio-temporal feature fusion layer includes a first adaptive spatio-temporal attention mechanism module, a first normalization layer, and a third relu activation function; wherein, the calculation formula adopted by the first adaptive spatio-temporal attention mechanism module is as follows:

[0064]

[0065] In formula (5), C D1 represents the number of channels of the matrix X D1 ; H D1 represents the height of the matrix X D1 ; W D1 represents the width of the matrix X D1 ; softmax represents the softmax activation function;

[0066] In step S3.3, the second depth feature extraction layer includes: a first 3×3 depthwise separable convolution, a first max pooling layer, a third sigmoid activation function, a second 3×3 depthwise separable convolution, a second max pooling layer, and a fourth sigmoid activation function connected in sequence; the second temporal feature extraction layer includes: a third bidirectional GRU recurrent neural network, a first 3×3 ordinary convolution, a fourth relu activation function, a fourth bidirectional GRU recurrent neural network, a second 3×3 ordinary convolution, and a fifth relu activation function connected in sequence;

[0067] In step S3.4, the second adaptive spatio-temporal feature fusion layer includes: a second adaptive spatio-temporal attention mechanism module, a second normalization layer, and a sixth relu activation function connected in sequence; wherein, the calculation formula adopted by this second adaptive spatio-temporal attention mechanism module is as follows:

[0068]

[0069] In formula (6), C D2 represents the number of channels of matrix X D2 ; H D2 represents the height of matrix X D2 ; W D2 represents the width of matrix X D2 ; softmax represents the softmax activation function;

[0070] In step S3.5, the global enhancement layer specifically includes: a 5×5 ordinary convolution, a 7×7 ordinary convolution, and a global average pooling layer connected in sequence; the local enhancement layer specifically includes: a 1×1 ordinary convolution, a third 3×3 ordinary convolution, and a third max pooling layer connected in sequence.

[0071] In an embodiment of the present invention, the calculation formula of the loss function used by the detection network module is as follows:

[0072]

[0073] In formula (7), L represents the number of samples; y i represents the true label of the i-th sample; p i represents the probability of the predicted label of the i-th sample; ∈ is a positive smoothing term.

[0074] In an embodiment of the present invention, step S5 specifically includes: inputting the signal data of different skin parts of a certain patient collected into the trained skin cancer detection model, and automatically obtaining the skin cancer detection result of this patient, so as to realize fast and accurate cancer detection of the skin of this patient.

[0075] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A skin cancer detection method based on deep learning, comprising: Step S1: collecting signal data from different skin parts of the patient; Step S2: designing a signal preprocessing module to preprocess the signal data; Step S3: sending the pre-processed signal data to the designed detection network module; Step S4: Design a loss function to obtain a skin cancer detection model; Step S5: Use the trained skin cancer detection model to perform skin cancer detection on the patient's skin.

2. The skin cancer detection method based on deep learning according to claim 1, characterized in that: The signal data includes: impedance data and phase angle data of different skin parts of the patient; and The patients include patients with normal skin, patients with benign skin and patients with malignant skin.

3. The skin cancer detection method based on deep learning according to claim 1, characterized in that: The signal preprocessing module includes a signal double-layer denoising processing module and a signal feature enhancement module; as well as The step S2 specifically includes: Step S2.1: performing double-layer denoising processing on two columns of data in the original signal matrix X corresponding to the signal data through the signal double-layer denoising processing module; wherein the two columns of data are impedance data and phase angle data respectively; Step S2.2: The matrix X after double-layer denoising m Send it to the signal feature enhancement module to obtain the matrix X e ; Step S2.3: Substitute the original signal matrix X and the matrix X e Concatenate to get the matrix X D .

4. The skin cancer detection method based on deep learning according to claim 3, characterized in that: Step S2.1 specifically includes: Step S2.11: Perform the first-layer denoising process on the two columns of data in the original signal matrix X corresponding to the signal data; the calculation formula corresponding to the first-layer denoising process is: In formula (1), X sij Represents the matrix X m The first layer of denoising value; X ij represents the data of the i-th row and j-th column of X; |min(X j )| represents the absolute value of the minimum value of the jth column; |max(X j )| represents the absolute value of the maximum value in the jth column; Step S2.12: Set X sij Perform the second layer of denoising. The calculation formula for the second layer of denoising is: In formula (2), X mij Represents the matrix X m The second layer denoising processing value; max means taking the larger of the two values; Step S2.2 specifically includes: mij Send it to the signal feature enhancement module to obtain the matrix X e ; The calculation formula corresponding to the signal feature enhancement module is: X eij =(softmax(w×relu(X mij +b)))×X mij +X mij (3) In formula (3), X eij Represents the matrix X e The value of ; softmax represents the softmax activation function; relu represents the relu activation function; w and b represent learnable parameters; and In step S2.3, the original signal matrix X and the matrix X e The calculation formula for splicing is: X D =reshape(X,X e )(4) In formula (4), reshape represents the horizontal splicing operation.

5. The skin cancer detection method based on deep learning according to claim 3, characterized in that: In step S3, the detection network module includes: a first deep feature extraction layer, a second deep feature extraction layer, a first temporal feature extraction layer, a second temporal feature extraction layer, a first adaptive spatiotemporal feature fusion layer, a second adaptive spatiotemporal feature fusion layer, a global enhancement layer, and a local enhancement layer; and The step S3 specifically includes: Step S3.1: Substituting the matrix X of the signal preprocessing module D The feature matrices output by the first depth feature extraction layer and the first time feature extraction layer are respectively sent to the first depth feature extraction layer and the first time feature extraction layer, and then the feature matrices output by the first depth feature extraction layer and the first time feature extraction layer are fused to obtain a matrix X D1 ; Step S3.2: Transform the matrix X D1 Send it to the first adaptive spatiotemporal feature fusion layer to obtain the matrix X TS1 ; Step S3.3: Transform the matrix X TS1 The feature matrices output by the second depth feature extraction layer and the second time feature extraction layer are respectively sent to the second depth feature extraction layer and the second time feature extraction layer, and then the feature matrices output by the second depth feature extraction layer and the second time feature extraction layer are fused to obtain the matrix X D2 ; Step S3.4: Transform the matrix X D2 Send it to the second adaptive spatiotemporal feature fusion layer to obtain the matrix X TS2 ; Step S3.5: Transform the matrix X TS2 respectively input into the global enhancement layer and the local enhancement layer to obtain feature matrices outputted by the global enhancement layer and the local enhancement layer respectively; Step S3.6: The feature matrices outputted by the global enhancement layer and the local enhancement layer are fused, and the fused feature matrices are sequentially passed through a fully connected layer and a softmax layer to obtain corresponding skin cancer detection results.

6. The skin cancer detection method based on deep learning according to claim 5, characterized in that: In step S3.1, the first depth feature extraction layer includes: a first 1×1 depthwise separable convolution, a first average pooling layer, a first sigmoid activation function, a second 1×1 depthwise separable convolution, a second average pooling layer, and a second sigmoid activation function connected in sequence; the first time feature extraction layer includes: a first bidirectional GRU recurrent neural network, a first 1×1 ordinary convolution, a first relu activation function, a second bidirectional GRU recurrent neural network, a second 1×1 ordinary convolution, and a second relu activation function connected in sequence; In step S3.2, the first adaptive spatiotemporal feature fusion layer includes a first adaptive spatiotemporal attention mechanism module, a first normalization layer and a third relu activation function; wherein the calculation formula adopted by the first adaptive spatiotemporal attention mechanism module is as follows: In formula (5), C D1 Represents the matrix X D1 The number of channels; H D1 Represents the matrix X D1 Height; W D1 Represents the matrix X D1 The width of ; softmax represents the softmax activation function; In step S3.3, the second depth feature extraction layer includes: a first 3×3 depth separable convolution, a first maximum pooling layer, a third sigmoid activation function, a second 3×3 depth separable convolution, a second maximum pooling layer, and a fourth sigmoid activation function connected in sequence; the second time feature extraction layer includes: a third bidirectional GRU recurrent neural network, a first 3×3 ordinary convolution, a fourth relu activation function, a fourth bidirectional GRU recurrent neural network, a second 3×3 ordinary convolution, and a fifth relu activation function connected in sequence; In step S3.4, the second adaptive spatiotemporal feature fusion layer includes: a second adaptive spatiotemporal attention mechanism module, a second normalization layer and a sixth relu activation function connected in sequence; wherein the calculation formula used by the second adaptive spatiotemporal attention mechanism module is as follows: In formula (6), C D2 Represents the matrix X D2 The number of channels; H D2 Represents the matrix X D2 Height; W D2 Represents the matrix X D2 The width of ; softmax represents the softmax activation function; In step S3.5, the global enhancement layer specifically includes: a 5×5 ordinary convolution, a 7×7 ordinary convolution, and a global average pooling layer connected in sequence; the local enhancement layer specifically includes: a 1×1 ordinary convolution, a third 3×3 ordinary convolution, and a third maximum pooling layer connected in sequence.

7. The skin cancer detection method based on deep learning according to claim 1, characterized in that: The calculation formula of the loss function used by the detection network module is as follows: In formula (7), L represents the number of samples; y i represents the true label of the i-th sample; p i Represents the probability of predicting the label of the i-th sample; ∈ is a positive smoothing term.

8. The skin cancer detection method based on deep learning according to claim 1, characterized in that: The step S5 specifically includes: inputting the collected signal data of different skin parts of a patient into a trained skin cancer detection model to automatically obtain the skin cancer detection result of the patient, so as to realize fast and accurate cancer detection on the patient's skin.

Images