1. The invention designs a novel verruca vulgaris auxiliary diagnosis and treatment method based on MultiResUnet, which is characterized by comprising the following steps: the method comprises the main steps of firstly obtaining the common wart pictures in a specific area, then carrying out data enhancement on the pictures to expand sample size, then making a binary label picture (a label of a lesion area and a background label) on the common wart pictures after data enhancement, then dividing the made data set into a training set, a verification set and a test set, wherein the ratio of the training set, the verification set and the test set is 6: 2, then inputting the processed data (comprising the training set and the verification set) into a MultiResUnet network for learning, carrying out multiple times of training, observing whether a loss function, model segmentation precision, MIoU (micro inertial measurement unit) is converged, whether the MIoU is improved or not after being subjected to super-parameter fine adjustment, and taking super parameters with more convergent loss functions, highest segmentation precision and the maximum MIoU value as final parameters of the model after training, and finally carrying out prediction and evaluation on the trained neural network model by using the test set.

2. The novel auxiliary diagnosis and treatment method for verruca vulgaris based on MultiResUnet as claimed in claim 1, wherein the method comprises the following steps: compared with the existing treatment method, the deep learning technology is adopted to assist the doctor in diagnosis and treatment instead of the doctor simply using naked eye diagnosis for diagnosis and treatment, and the method can reduce the occurrence of misdiagnosis caused by fatigue of the doctor, thereby reducing unnecessary harm to the patient.

3. The novel auxiliary diagnosis and treatment method for verruca vulgaris based on MultiResUnet as claimed in claim 2, wherein: compared with UNET, MultiResUnet replaces convolution of two 3X3 in UNET by 3X3, 7X7 convolution operation is combined with 5X5 convolution operation in parallel, Res Path is used for replacing simple Skip Connection in traditional UNET, and parameter quantity is less than that of UNET but efficiency is higher than that of UNET.

4. The novel auxiliary diagnosis and treatment method for verruca vulgaris based on MultiResUnet as claimed in claim 3, wherein: according to the invention, the convolutional layer activation function in the Multi _ Res _ Block module in the MultiResUnet is changed into PReLU to replace ReLU, and the convolutional layer activation function in the Res _ Path module adopts ReLU, so that experiments show that the result of using PReLU + ReLU is better than that of using ReLU alone. And PReLU can be used to solve the problem of neuronal necrosis brought about by ReLU.

5. The novel auxiliary diagnosis and treatment method for verruca vulgaris based on MultiResUnet as claimed in claim 4, wherein: multiple experiments show that under the condition of the same data set and the same super parameter, the UNET and the MultiResUnet are trained for the same cycle iteration times, and the result shows that the Loss Function of the MultiResUnet network is fast in convergence, good in image segmentation effect and high in MIoU value.

Background

With the continuous development of the deep learning field in recent years, the technology of the deep learning in the medical image processing field is more and more mature. The method is widely applied to the field of medical image segmentation. The pixel-level classification of images is performed by the full volumetric Network (FCN) proposed by Long et al, thereby solving the semantic-level image segmentation problem. Unlike the classic CNN, which uses a fully-connected layer to obtain fixed-length feature vectors for classification in convolutional layers, the FCN can accept input images of any size, and uses a deconvolution layer to up-sample the feature map of the last convolutional layer to restore it to the same size as the input image, thereby generating a prediction for each pixel and simultaneously retaining spatial information in the original input image. Thus, FCN becomes a corner stone for deep learning to solve the segmentation problem. UNET is a network improved on the basis of FCN, and mainly includes a down-sampling part and an up-sampling part. The down-sampling part is mainly responsible for image feature extraction, and the up-sampling part performs feature fusion and restores the size of the image. UNET employs a completely different feature fusion approach: stitching, UNET, employs stitching features together in the Channel dimension to form thicker features. And the corresponding points used in FCN fusion add up and do not form thicker features. And therefore, the application of UNET in the field of image segmentation is wider than that of FCN. However, the classical UNET architecture seems to be lacking in some respects, and a multiresonet network structure is born. The method improves the UNET structure, replaces convolution of two 3X3 in the UNET by 3X3, combines 7X7 convolution operation and 5X5 convolution operation in parallel, replaces simple Skip Connection in the traditional UNET by Res Path, and has less parameter quantity but higher efficiency than the UNET.

The invention utilizes the MultiResUnet network architecture to assist a doctor in diagnosing the position of the common wart of a patient and carrying out the next treatment, thereby improving the misdiagnosis condition of the doctor and greatly improving the working efficiency of the doctor.

Disclosure of Invention

The invention designs a novel common wart auxiliary diagnosis and treatment method based on MultiResUnet, which is mainly used for solving the problems that the shapes of partial common warts are small and not obvious and the number of the common warts is variable due to infectivity, a doctor judges the position of cryotherapy by naked eyes only with difficulty and time consumption, and when a patient to be examined is large in quantity, the doctor judges the position of the common wart by mistake due to fatigue so as to hurt other unrelated skin tissues and cause unnecessary damage to the patient. In order to achieve the purpose, the adopted processing method is that a multiResUnet network is trained in a large amount by using an ordinary wart picture data set of a specific area (a relatively flat area, such as a palm, a back of a hand, a sole and the like), a segmentation model with the converged loss function, the highest precision and the maximum MIoU value is finally obtained through fine adjustment of hyper-parameters and the like, and an ordinary wart picture of the specific area is input into the network, so that a rough area with the bottom of the ordinary wart positioned on the skin surface can be segmented.

The method comprises the main steps of firstly obtaining the common wart pictures in a specific area, then carrying out data enhancement on the pictures to expand the sample size, and then making a binary label picture (a label of a lesion area and a background label) on the common wart pictures after the data enhancement. Then, the manufactured data set is divided into a training set, a verification set and a test set, and the proportion of the training set, the verification set and the test set is 6: 2. And inputting the processed data (including a training set and a verification set) into a MultiResUnet network for learning. And training for multiple times, observing whether the loss function and the model segmentation precision are converged or not and whether the MIoU value is improved or not after the hyper-parameter fine adjustment, and adopting the hyper-parameters with the more converged loss function, the highest segmentation precision and the largest MIoU value as final parameters of the model after training. And finally, carrying out prediction evaluation on the trained neural network model by using a test set.

Different from the existing treatment method, the invention has the beneficial effects that:

1. compared with the existing treatment method, the deep learning technology is adopted to assist the doctor in diagnosis and treatment instead of the doctor simply using naked eye diagnosis for diagnosis and treatment, and the method can reduce the occurrence of misdiagnosis caused by fatigue of the doctor, thereby reducing unnecessary harm to the patient.

2. Compared with UNET, MultiResUnet replaces convolution of two 3X3 in UNET by 3X3, 7X7 convolution operation is combined with 5X5 convolution operation in parallel, Res Path is used for replacing simple Skip Connection in traditional UNET, and parameter quantity is less than that of UNET but efficiency is higher than that of UNET.

3. According to the invention, the convolutional layer activation function in the Multi _ Res _ Block module in the MultiResUnet is changed into PReLU to replace ReLU, and the convolutional layer activation function in the Res _ Path module adopts ReLU, so that experiments show that the result of using PReLU + ReLU is better than that of using ReLU alone. And PReLU can be used to solve the problem of neuronal necrosis brought about by ReLU.

4. Multiple experiments show that under the condition of the same data set and the same super parameter, the UNET and the MultiResUnet are trained for the same cycle iteration times, and the result shows that the Loss Function of the MultiResUnet network is fast in convergence, good in image segmentation effect and high in MIoU value.

Drawings

The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.

Fig. 1 is a flow chart of the present invention for segmenting verruca vulgaris in a particular region using a multiresonet network;

fig. 2 is a Res Path structure diagram in the multiresonet network employed in the present invention.

Detailed Description

The following detailed description of embodiments of the invention refers to the accompanying drawings.

Please refer to fig. 1:

a novel verruca vulgaris auxiliary diagnosis and treatment method based on MultiResUnet specifically comprises the following five steps:

the method comprises the following steps: acquiring a verruca vulgaris picture of a specific area, and performing data enhancement, wherein the main modes comprise: turning over, rotating, blurring, adjusting brightness and the like, then making a binary label image (a label of a lesion area and a background label) on the verruca vulgaris image after data enhancement, and entering a second step;

step two: and dividing the processed data set. Dividing the data set into a training set, a verification set and a test set, wherein the ratio of the training set to the verification set to the test set is 6: 2, and entering the third step;

step three: a MultiResUnet network is adopted, and a convolution layer ReLU activation function in a MultiRes _ Block module is replaced by a PReLU to solve the problem of neuron necrosis caused by the ReLU;

step four: inputting the divided data sets (including a training set and a verification set) into a MultiResUnet network, learning in a mode of training and verifying, wherein a loss function adopts a cross entropy function, an optimizer uses Adam, and a segmentation evaluation index adopts a Mean cross over ratio (MIoU, a calculation formula:wherein k +1 represents the number of categories (including empty categories), i represents the true value, j represents the predicted value, and Pij represents that i is predicted to be j. Or may be deformed intoWherein k +1 represents the number of categories (including empty categories), TP represents a true positive case, FP represents a false true case, FN represents a false negative case, and TN represents a true negative case), and training is performed for multiple times, and whether a loss function and model segmentation precision are converged, improved or increased after the MIoU is subjected to hyper-parameter fine tuning is observed, and after the training is completed, hyper-parameters with the loss function more converged, the segmentation precision highest and the MIoU value largest are adopted as final parameters of the model, and the method enters step five;

step five: and predicting the trained neural network model by using the test set, and evaluating the generalization ability of the model.

Please refer to fig. 2:

fig. 2 is a Res Path structure diagram in the multiresonet network employed in the present invention. The method is different from the method that the feature diagram obtained by the encoder is directly subjected to feature fusion with the feature diagram output by the decoder in UNET, Res Path performs convolution operations of 3x3 and 1x1 on the feature diagram obtained by the encoder, residual errors of the feature diagram obtained by the convolution are connected, and then the final result is subjected to feature fusion with the feature diagram output by the decoder.