CN115005768A

CN115005768A - Skin disease picture classification method, device, product and storage medium

Info

Publication number: CN115005768A
Application number: CN202210362657.7A
Authority: CN
Inventors: 高丽华; 张博; 康健; 管昊瑜; 蒋佳伶; 陈菲雪; 毕英琦; 郝文龙; 谢琪; 黄小丁; 周大维
Original assignee: Third Xiangya Hospital of Central South University
Current assignee: Third Xiangya Hospital of Central South University
Priority date: 2022-04-08
Filing date: 2022-04-08
Publication date: 2022-09-06

Abstract

The invention discloses a skin disease picture classification method, a device, a product and a storage medium, which are used for constructing a basic object of a skin disease RCM image database; RCM images of different skin diseases are filled into the corresponding data sets. Constructing a CNN-LSTM-based skin disease image classification system for identifying RCM images; training and optimizing the classification system until the classification accuracy reaches the expectation. The invention greatly improves the classification efficiency and accuracy of the skin disease images.

Description

Skin disease picture classification method, device, product and storage medium

Technical Field

The invention relates to an image processing technology, in particular to a method, a device, a product and a storage medium for classifying pictures of skin diseases.

Background

In recent years, the application of the Reflection Confocal Microscope (RCM), abbreviated as skin CT, in dermatology is becoming more and more widespread, because it can realize the real-time imaging of the epidermis and the dermis superficial cell level in vivo and non-invasively. In China, many advanced imaging devices are idle for a long time due to the lack of doctors capable of diagnosing examination results, which is a great waste of medical resources. Therefore, in order to solve the above problems, it is important to search for a new method and open a new way for the whole patient with a huge amount of patients and advanced technology with a change of day and night.

At present, most of the existing artificial intelligence skin disease image classification systems are mainly capable of performing identification and classification tasks of single medical images, and most of the existing artificial intelligence skin disease image classification systems adopt mature CNN as a framework, for example, the prior art 1 (application number: 201910334631.X, publication number: CN 110895968A) discloses an artificial intelligence medical image automatic diagnosis system and method, and the segmentation and classification of medical microscope images, particularly gynecological micro-ecological microscope images, are realized through the optimal AI classification model and the optimal AI semantic segmentation model. However, when a series of consecutive medical images such as RCM images and MRI images are related to each other, the conventional CNN framework cannot perform an effective comprehensive analysis on the images. The skin diseases are complex and variable, so that the surface image of the skin damage cannot completely show the pathological features of the skin diseases. In order to ensure complete extraction of lesion features of skin lesions, classification of skin diseases should be based on multiple layers of skin images of epidermis, acanthosphere, stratum granulosum and the like of a patient. The existing artificial intelligent skin disease image classification system lacks the comprehensive analysis capability of multiple associated images and can only extract and classify the features of a single skin image, so that the accuracy of an intelligent classification result obtained based on a single skin damage surface image is low, and the reliability is weak.

Disclosure of Invention

The invention aims to solve the technical problem that the prior art is insufficient, and provides a method, a device, a product and a storage medium for classifying skin disease images, which can be used for identifying a plurality of images with different depths and improving the accuracy rate of classifying the skin disease images.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows: a skin disease picture classification method comprises the following steps:

s1, constructing a skin disease RCM image database;

s2, dividing the RCM image database of the skin diseases into a training set and a test set;

s3, training a classification model based on CNN-LSTM by using a training set to obtain a skin disease image classification system; the classification model of CNN-LSTM comprises an encoder and a decoder; the encoder adopts a CNN network, and the decoder adopts an LSTM network;

s4, the RCM images to be processed are sorted according to the skin depth, and the sorted image sequences are input into the skin disease image classification system to obtain a classification result.

The skin disease picture classification system for recognizing RCM images based on CNN-LSTM can integrally analyze a plurality of continuously related RCM images by combining the convolutional neural network and the long-term and short-term memory network, and overcomes the defect that the traditional skin disease auxiliary classification system can only recognize a single image.

In order to further improve the image classification accuracy, after step S3 and before step S4, the method further includes: and testing the skin disease image classification system by using a test set, if the test result of the skin disease image classification system fails, expanding the skin disease RCM image database to obtain an expanded training set, and training the classification model based on the CNN-LSTM by using the expanded training set.

The specific implementation process of step S1 includes:

1) collecting RCM skin disease images, and numbering the collected RCM images;

2) constructing a basic object of a skin disease RCM image database, wherein the basic object comprises a plurality of tables, and the tables comprise RCM images for storing different types, image shooting depths and image numbers;

3) and filling the collected RCM skin disease images into a corresponding table according to the image shooting depth and the image types to obtain an RCM image database.

The RCM image is classified according to the type and the skin depth, and the classification precision is further improved.

In step S3, the CNN network performs an operation on the input RCM image, the operation including the following steps: respectively inputting RCM images of an epidermal layer cornification layer, an epidermal layer transparent layer, an epidermal granular layer, an epidermal layer hair growth layer, a dermal papillary layer and a dermal reticular layer of RCM images with the same image number into a CNN network at the moments of t1, t2, t3, t4, t5 and t6, converting each layer of image into a two-dimensional array, and performing convolution operation on the two-dimensional array to obtain a feature vector diagram corresponding to the two-dimensional array; and inputting the feature vector diagram into an activation function and a pooling layer in sequence to obtain the standardized RCM image features. According to the invention, the images are sequentially input into the CNN system from shallow to deep according to skin layers, the shooting depth attribute of the images is coded by time, and the shooting depth problem which cannot be processed by the LSTM system is converted into the time sequence problem which can be processed by the LSTM system, so that the CNN-LSTM system can analyze the images based on the shooting depths of different RCM images and comprehensively analyze a plurality of images at different depths, and the problems of low accuracy and poor reliability of an intelligent classification result obtained based on a pure skin damage surface image in the prior art are solved.

In step S3, the LSTM network performs operations including the steps of: and sequentially inputting the normalized RCM image feature sequences S1, S2, S3, S4, S5 and S6 output by the CNN network into the LSTM at time points t1 ', t 2', t3 ', t 4', t5 'and t 6', and extracting feature information of the image features acquired at each moment by a hidden layer of the LSTM network and integrating the image information with image information of the rest skin layers at the historical moment.

In step S3, the error equation E of the CNN-LSTM classification model is:

wherein z is _t A classification result given by a CNN-LSTM classification model after a group of RCM images are input; y is _t Correct classification results corresponding to a group of input RCM images; p is the index of the training sample in the LSTM grid, the training sample is the RCM image of the cornified epidermis layer, the stratum lucidum layer, the granular epidermis layer, the germinal epidermis layer, the papillary dermis layer, and the reticular dermis layer of a set of RCM images with the same image number, and the images are respectively input into the system at the time points t1, t2, t3, t4, t5, and t 6. z is a radical of _t A classification result is given by a CNN-LSTM classification model after the RCM image is input at the time t; y is _t Correct classification results corresponding to a group of input RCM images; (ii) a T is a constant (set to 6 in the present invention).

The error equation of the invention can accurately judge the accuracy of the system when the system is trained.

In step S3, the classification model of CNN-LSTM is trained using a gradient descent method.

The invention also provides a computer device, comprising a memory, a processor and a computer program stored on the memory; the processor executes the computer program to implement the steps of the above-described method of the present invention.

The present invention also provides a computer program product comprising a computer program/instructions; which when executed by a processor, performs the steps of the above-described method of the present invention.

The present invention also provides a computer readable storage medium having stored thereon a computer program/instructions; which when executed by a processor implement the steps of the above-described method of the present invention.

Compared with the prior art, the invention has the following beneficial effects:

1. the invention relates to a CNN-LSTM-based dermatosis picture classification system for identifying RCM images, which applies core technologies of a convolutional neural network and a long-term and short-term memory network, can integrally analyze a plurality of continuously related RCM images and makes up the defect that the traditional dermatosis auxiliary classification system can only identify a single image.

2. The method has the characteristics of high classification speed and high classification accuracy, can effectively assist doctors in classifying the skin diseases, and is favorable for relieving the problem that the disease condition of a patient is delayed when the proportion of current dermatologists and patients is unbalanced and the disease condition of the patient is not diagnosed in time;

3. the invention applies CNN-LSTM to classification tasks of RCM and skin CT, and the images have the characteristics that: images do not exist in isolation, each group of images contains different information of skin layers from light to deep, if CNN is applied to classification tasks, only images of a certain layer can be classified, and image information of other layers is not fully utilized. The CNN-LSTM provided by the invention can be used for classifying images of all levels, has the capability of correlation analysis among images of different levels, and overcomes the defect of low data utilization rate when the traditional CNN is applied to RCM and skin CT classification tasks.

4. The invention also has memorability, can classify a group of images input in sequence, can utilize the picture information of each skin layer of the same lesion part, has more available information compared with the analysis of a single picture, and has higher classification accuracy.

Drawings

FIG. 1 is a flow chart of a method of an embodiment of the present invention;

FIG. 2 is a build process of an embodiment of the invention;

FIG. 3 is a process of CNN-LSTM processing of an image;

FIG. 4 is a process of CNN extracting image feature information;

FIG. 5 is a process of LSTM processing time-dependent information;

FIG. 6 is a process of CNN-LSTM training;

FIG. 7 is a specific process of the system identifying RCM images of skin diseases;

FIG. 8 is a process of analysis processing of a continuous input image by the CNN-LSTM system;

FIG. 9 is a chart of the result of the comparison of the accuracy of the classification system of skin diseases based on CNN and the system of the present invention;

fig. 10 is a comparison of the completeness of the CNN-based skin disease picture classification system and the system of the present invention.

Detailed Description

The invention provides a skin disease picture classification system based on a convolutional neural network and a long-short term memory network. A group of RCM influences are sequentially input into a CNN-LSTM system according to a shooting depth sequence, characteristic information of each image is extracted through the CNN, and a time-dependent sequence output by the CNN is processed through the LSTM, so that comprehensive analysis of skin image information at different depths is realized, the defect that an existing image classification system can only recognize a single skin image is overcome, and the accuracy and reliability of skin disease image classification are further improved.

The technical solution of the present application will be described in detail below with specific examples. Example 1 is a construction process for system function implementation. Example 2 a specific procedure for the system to identify skin disorders.

Example 1

Referring to fig. 1, the system function is realized mainly depending on two plates of the skin disease RCM image database and the CNN-LSTM-based skin disease image classification system. And training the CNN-LSTM frame through the ordered classified images called from the database, so as to ensure the realization of functions.

Referring to fig. 2, the system is constructed mainly by the following parts:

s101, constructing a skin disease RCM image database.

Specifically, the step of constructing the skin disease RCM image database includes:

the data of the high-quality skin disease RCM image and skin disease RCM image database is collected from a hospital electronic disease case library, and the skin disease RCM image database is constructed by collecting high-quality skin disease histories in the hospital electronic disease case library.

The collected RCM images are numbered. The RCM images of the same patient and the same examination are generally a set of images including RCM images of different depth tomograms, and each image set is numbered. All RCM images in the unified image group share the same picture number.

Constructing a skin disease RCM image database basic object, wherein the basic object comprises but is not limited to a table and a storage process. The table is used for storing different types of RCM images and related information, including but not limited to epidermis layer cuticle layer vitiligo image table, epidermis layer diaphragma vitiligo image table, epidermis layer granular layer vitiligo image table, epidermis hair growth layer vitiligo image table, dermis layer papillary layer vitiligo image table, dermis layer reticular layer chloasma image table and the like. The attributes of each table include, but are not limited to, RCM image, skin disease type, image capture depth, picture number. The storage process includes, but is not limited to, having the function of extracting the same set of RCM images of the same picture number.

And filling the data of the dermatological disease cases. Storing the collected RCM images of the high-quality skin disease cases into a corresponding table according to the scanning depth and the skin disease types of the RCM images, and filling other related information required by the data set;

the resulting patient case data set is obtained. When the number of collected medical records is more than 5000, dividing the data of the dermatological disease cases into a training set and a testing set according to the ratio of 4: 1.

S102, constructing a CNN-LSTM model

FIG. 3 shows the processing of the image through the CNN-LSTM model.

The specific steps for constructing the CNN-LSTM model comprise:

and (5) performing RCM image feature extraction by using the CNN. Referring to fig. 4, the CNN feature extraction process is as follows. RCM images of the cuticular layer, the epidermal clear layer, the granular layer, the epidermal germinal layer, the dermal papillary layer, and the dermal reticular layer having the same picture number are defined as P1, P2, P3, P4, P5, and P6. P1, P2, P3, P4, P5, and P6 are input to CNN at times t1, t2, t3, t4, t5, and t6, respectively. For a single picture, the picture information is converted into a two-dimensional array through a CNN input layer, convolution operation is carried out through a convolution layer, namely, a convolution kernel is scanned on a skin image on each channel according to a proper size, and the image is converted into a high-dimensional feature vector diagram. At this stage, the feature information of the RCM picture is extracted and enhanced. And activating the enhanced feature information through a Relu function, inputting the enhanced feature information into a pooling layer for pooling processing and dimension reduction by compression, and obtaining the main features of the normalized RCM image with less data volume.

And carrying out sequence data analysis by using the LSTM to obtain a classification result. Referring to fig. 5, the LSTM data analysis process is as follows. After the CNN sequentially performs feature extraction on P1, P2, P3, P4, P5 and P6, time-dependent sequences S1, S2, S3, S4, S5 and S6 are obtained, and are input to the LSTM as input at each time point at t1 ', t 2', t3 ', t 4', t5 'and t 6'. The hidden layer of the LSTM processes the RCM image characteristic information which is transmitted by the CNN at a certain moment and represents a certain skin depth, meanwhile, the long-term memory storage is realized by means of the updating of the cell state, the RCM image characteristics of the layer analyzed at the moment are collected and integrated with the RCM image characteristic information of other skin layers analyzed at all the moments before, finally, the RCM image characteristics of all layers are integrated at the last moment, a comprehensive analysis result is obtained, and the comprehensive analysis result is output through the output layer. Taking the time t4 'as an example, the high-dimensional vector representing the RCM image features of the epidermal growth layer, P4, is input to the LSTM, the LSTM processes the high-dimensional vector, and then updates the cell state, and the processing information is stored in a manner integrated with the processing information of the high-dimensional vector representing the RCM image features of the epidermal cornification layer, the epidermal transparent layer, and the epidermal granular layer, which is input at the times t 1', t2 ', and t 3'.

The CNN-LSTM model is trained using a training set. FIG. 6 shows the process of CNN-LSTM training. And using the training set of S201 for training the model. The error is reversely propagated through the BPTT algorithm, so that the classification function of the system is realized, and the classification accuracy of the system is continuously improved on the basis.

The error equation defining the entire network is:

wherein zt is the system output corresponding to the input data, namely the classification result given by the system after a group of RCM images are input; yt is the actual output corresponding to the input data, namely the correct classification result corresponding to the input group of RCM images; p is the index of the training samples in the LSTM grid.

In the process of error propagation in each CNN-LSTM layer, calculating the gradient of each weight and bias, updating each parameter of the system by a gradient descent method, minimizing an error equation, obtaining an optimal system structure, and improving the classification accuracy of the system.

And carrying out a plurality of training iterations, and finally, after all the training sample classification results are processed according to the error equation, enabling the total error E to fall within the expected minimum value, thereby completing the training process of the system.

For example, a group of RCM image groups diagnosed as vitiligo are selected, sequentially input into the CNN-LSTM model according to the image shooting depth, and finally obtained classification results generate errors if the classification results are not vitiligo after the CNN and LSTM layers are processed. And (4) performing reverse error propagation through a BPTT algorithm, and adjusting the weight and the bias gradient of each CNN-LSTM layer according to the error. After a large number of different image groups are trained repeatedly, all parameters of the system are continuously readjusted and updated, so that errors are minimized, and the accuracy reaches the highest.

S103, evaluating classification function of CNN-LSTM

Specifically, the system classification accuracy is tested using a test set.

The system adopts recognized classification evaluation indexes to evaluate the classification result of the skin mirror image, and comprises the following steps:

accuracy (Accuracy):

precision (Precision):

recall (Recall):

where Nr represents the number of correctly sorted samples, Na represents the total number of samples, tp represents true positives, fp represents false positives, and fn represents false negatives.

When A, p and R are all more than 95%, the system performance test is over-closed, and the system can be put into use; otherwise, continuing to collect the high-quality medical records, expanding the RCM image database of the skin diseases, and using the expanded data to train the system again.

Example 2

Fig. 7 shows a specific process of the system to identify the effects of a skin disorder RCM.

S201, inputting the image into the CNN-LSTM according to requirements.

Specifically, RCM images of the epidermis cornified layer, the epidermis diaphanous layer, the epidermis granular layer, the epidermis germinal layer, the dermis papillary layer, and the dermis reticular layer photographed by the patient are defined as a1, a2, A3, a4, and a5, and a6 is input to the CNN-LSTM system at times t1, t2, t3, t4, t5, and t6, respectively.

S202, the CNN-LSTM system analyzes and processes the continuous input images.

Referring to fig. 8, the process of analyzing and processing the continuous input images by the CNN-LSTM system is as follows:

at time t1, when a1, namely the epidermis layer angulation layer RCM image input system, is available, the CNN performs feature extraction on a1, and the obtained high-dimensional vector representing the features of the epidermis layer angulation layer RCM image is input to the LSTM at time t 1', and a processing result R1, namely a classification result based on the epidermis layer angulation layer RCM image, is obtained.

At time t2, when a2, i.e., the epidermis transparent layer RCM image input system, the CNN performs feature extraction on a2, and the obtained high latitude vector representing the image feature of the epidermis transparent layer RCM image is input to the LSTM at time t 2', and the LSTM updates the cell state, and obtains a processing result R1+ R2 — a classification result based on the epidermis layer angulation layer and the epidermis transparent layer RCM image.

At time t3, when A3, namely the system for inputting the RCM image of the epidermis granular layer, CNN performs feature extraction on A3, and the obtained high-dimensional vector representing the RCM image of the epidermis granular layer is input to LSTM at time t 3', and the LSTM updates the cell state to obtain a processing result R1+ R2+ R3, which is a classification result based on the epidermis cornified layer, the epidermis transparent layer and the RCM image of the epidermis granular layer.

At time t4, there is a4, i.e., the skin layer growth layer RCM image input system, the CNN performs feature extraction on a4, and the obtained high-dimensional vector representing the features of the skin layer growth layer RCM image is input to the LSTM at time t 4', and the LSTM updates the cell state to obtain a processing result R1+ R2+ R3+ R4 — a result of classification based on the skin layer cuticle layer, skin layer transparency layer, skin layer granulation layer, and skin layer growth layer RCM images.

At time t5, when a5, namely an RCM image input system of a papillary layer of a dermis, is available, the CNN performs feature extraction on a5, the obtained high-dimensional vector representing the RCM image features of the papillary layer of the dermis is input into the LSTM at time t 5', the LSTM updates the cell state, and a processing result R1+ R2+ R3+ R4+ R5 is obtained, wherein the processing result is based on the classification results of RCM images of an epidermis cornified layer, an epidermis transparent layer, an epidermis granular layer, an epidermis germinal layer and the papillary layer of the dermis.

At time t6, when a6, namely the RCM image input system of the reticular layer of the dermis, is available, the CNN performs feature extraction on a6, the obtained high-dimensional vector representing the RCM image features of the reticular layer of the dermis is input into the LSTM at time t 6', the LSTM updates the cell state, and a processing result R1+ R2+ R3+ R4+ R5+ R6 is obtained, wherein the processing result is based on the classification results of RCM images of the cornified layer of the epidermis, the stratum lucidum layer of the epidermis, the stratum granulosum layer of the epidermis, the germinal layer of the epidermis, the papillary layer of the dermis and the reticular layer of the dermis. At this point, the system input is complete.

Further, if there are a plurality of images missing in the RCM image of the epidermis transparent layer, the epidermis granular layer, the epidermis hair growth layer, the dermis papillary layer, and the dermis reticular layer, a blank picture is input at the time of inputting the image of the layer.

Further, "+" in the processing result of the LSTM indicates that the memory integration of the results at the time and all previous times is not simply added.

S203, the CNN-LSTM system finishes the classification of the continuous images to obtain a classification result.

After analyzing and processing the continuous input images, the CNN-LSTM system processes the final processing result, converts the machine language into the human language and feeds back the classification result to the user.

For example, for a1 showing hyperkeratosis or parakeratosis, A3 showing thinning or disappearance of the stratum granulosum, a4 showing thickening of the stratum spinosum, regular downward extension of the epicutaneous process, significant reduction of pigment in the basal cell layer, a5 showing a set of RCM images of vascular tortuosity in the dermal papilla layer, and mild to moderate inflammatory cell infiltration around the blood vessel, the final output classification of the system is vitiligo.

250 RCM image classifications were tested before and after access to the LSTM. Selecting RCM images of 25 patients, 10 patients, selecting one RCM image from the RCM images of each patient randomly by CNN for classification, and sequentially classifying 10 RCM images from shallow to deep levels of each patient by CNN-LSTM.

Example 3

The CNN-LSTM-based RCM dermatological image classification system of example 1 was evaluated, with the following specific evaluation procedures:

1 dermatologists, attending physicians and inpatients are randomly extracted from a certain hospital, the classification accuracy of the dermatologists, the attending physicians and the inpatients is compared with the classification system of the skin disease pictures based on the CNN and the classification system of the skin disease pictures based on the CNN in the

embodiment

1, 10 dermatologists are extracted to form an expert committee, and the expert committee is responsible for making standard answers of the comparison experiment.

The above-mentioned dermatologists, attending physicians, and hospitalizers are required to classify 400 cases within 4 hours without depending on other related information.

The following data were counted every 1h for the dermatologist, attending physician, hospitalized physician:

accuracy (Accuracy):

wherein, N _r Number of samples representing correct classification, N _c Indicating the number of samples completed at the current time.

Degree of completion (completensiss):

wherein, N _c Indicating the number of samples completed, N _a Represents the total number of samples.

Accuracy and completeness contrast maps of the dermatologist, attending physician, resident physician, CNN-based dermatologic image classification system and the system of the present invention were obtained as shown in fig. 9 and 10. As can be seen from FIG. 9, in 1H, the dermatologists, the main doctors and the system have higher accuracy, and the accuracy reaches more than 89%; with the increase of time, the accuracy of dermatologists, treating doctors and hospitalized doctors begins to decline, while the accuracy of the skin disease image classification system based on CNN and the accuracy of the system of the invention gradually rise, and the accuracy of the system of the invention reaches more than 90 percent, so that the system of the invention has better performance compared with the skin disease image classification system based on dermatologists, treating doctors, hospitalized doctors and CNN on the basis of the measurement index of the accuracy. As can be seen from fig. 10, the test completion of the system of the present invention and the CNN-based skin disease image classification system was maintained at a 100% level during the experiment; the completion degree of dermatologists, attending physicians and hospitalized physicians is reduced along with the increase of time, wherein the completion degree of the dermatologists is the highest, but the highest completion degree is only 30 percent and is far lower than that of the system, so that the system greatly improves the classification efficiency.

Claims

1. A skin disease picture classification method is characterized by comprising the following steps:

s1, constructing a skin disease RCM image database;

s2, dividing the skin disease RCM image database into a training set and a testing set;

2. The method for classifying skin diseases images according to claim 1, wherein after step S3 and before step S4, the method further comprises: and testing the skin disease image classification system by using a test set, if the test result of the skin disease image classification system fails, expanding the skin disease RCM image database to obtain an expanded training set, and training the classification model based on the CNN-LSTM by using the expanded training set.

3. The method for classifying skin diseases picture according to claim 1, wherein the step S1 is implemented by the following steps:

1) collecting RCM skin disease images, and numbering the collected RCM images;

2) constructing a basic object of a dermatosis RCM image database, wherein the basic object comprises a plurality of tables, and the plurality of tables comprise RCM images for storing different types, image shooting depths and image numbers;

4. The method for classifying skin diseases picture according to claim 1, wherein in step S3, the CNN network performs the following steps on the input RCM image: respectively inputting RCM images of an epidermal layer cutinization layer, an epidermal layer transparent layer, an epidermal granular layer, an epidermal layer hair growth layer, a dermal papilla layer and a dermal reticular layer of RCM images with the same image number into a CNN network at the time of t1, t2, t3, t4, t5 and t6, converting the images of all layers into a two-dimensional array, and performing convolution operation on the two-dimensional array to obtain a feature vector diagram corresponding to the two-dimensional array; and inputting the feature vector diagram into an activation function and a pooling layer in sequence to obtain the standardized RCM image features.

5. The method for classifying skin diseases picture according to claim 4, wherein in step S3, the LSTM network performs the operation comprising the steps of: and sequentially inputting the normalized RCM image feature sequences S1, S2, S3, S4, S5 and S6 output by the CNN network into the LSTM at time points t1 ', t 2', t3 ', t 4', t5 'and t 6', respectively, extracting feature information of the image features acquired at each moment by a hidden layer of the LSTM network, and integrating the image information with image information of other skin layers at historical moments.

6. The method for classifying skin diseases image according to claim 1, wherein in step S3, the error equation E of the classification model of CNN-LSTM is:

wherein p is the index of the training sample in the LSTM grid, the training sample is a group of RCM images of the cuticular layer, the papillary layer and the reticular layer of the cuticular layer of the RCM images with the same image number, the images are respectively input into the classification model of the CNN-LSTM at the time points of t1, t2, t3, t4, t5 and t6, and z is the index of the training sample in the LSTM grid _t Score given by classification model of CNN-LSTM after RCM image input at time tClass results; y is _t Inputting a correct classification result corresponding to a group of RCM images; t is a constant.

7. The method for classifying skin diseases according to claim 1, wherein in step S3, the classification model of CNN-LSTM is trained by gradient descent method.

8. A computer apparatus comprising a memory, a processor and a computer program stored on the memory; characterized in that the processor executes the computer program to implement the steps of the method according to one of claims 1 to 6.

9. A computer program product comprising a computer program/instructions; characterized in that the computer program/instructions, when executed by a processor, performs the steps of the method according to one of claims 1 to 6.

10. A computer readable storage medium having stored thereon a computer program/instructions; characterized in that the computer program/instructions, when executed by a processor, implement the steps of the method of one of claims 1 to 6.