Research On Dermoscopy Images Segmentation And Classification Method Based On Convolutional Neural Network

Skin cancer and various pigmented dermatoses are seriously threatening human health.At present,the diagnosis of pigmented skin disease mainly realized by observing and analyzing the lesion features shown in the dermoscopy image.The dermoscopy image is a medical image obtained by non-invasive microscopic imaging technology,which can clearly show the characteristics of the pigmented dermatoses lesion.However,due to the small difference of lesions in different cases,it has become very difficult for doctors to analyze the type of lesions by observing the dermoscopy image with naked eyes.In order to achieve effective treatment,the demand for computer aided diagnosis systems for dermoscopy image has increased.Computer aided diagnosis can alleviate the pressure of doctors and help improve the efficiency and accuracy of diagnosis.The computer-aided diagnosis system for dermoscopy images mainly includes two tasks: dermoscopy image segmentation and classification.The purpose of dermoscopy image segmentation is to determine the contour and size of the lesion to ensure the accuracy of surgical resection.The mainstream dermoscopy image segmentation methods include region-based segmentation,threshold-based segmentation,cluster-based segmentation,and supervised learning methods.These methods are influenced by subjective factors and impurities such as hair and blisters in images,and the segmentation effect is not ideal.Although the convolutional neural network can accomplish the semantic segmentation task of natural pictures well,it is not mature in the field of dermoscopy image segmentation,and the segmentation effect still has much room for improvement.For the dermoscopy image classification task,many studies have manually extracted the features of the lesions in the dermoscopy image and then combined with computer algorithms to classify the features to determine the type of lesions.This method requires manual participation,which is difficult and influenced by subjective factors.The convolutional neural network in the classification of dermoscopic images focuses on the use of a single network for melanoma recognition.There still have a lot of room for improvement in the accuracy of various types of pigmented dermatoses classification.In view of the shortcomings of the above mainstream methods,this paper presents a method based on convolutional neural network to accurately and automatically realize the segmentation and classification task.The main work of this paper is as follows:1.A dermoscopy image segmentation and classification model based on convolution neural network is presented.The model includes dermoscopy image acquisition module,dermoscopy image denoise module,image segmentation module,image classification module and test module.Each module realizes related functions such as dermoscopy image collection,dermoscopy image color difference noise removal,image segmentation,and image classification.2.In order to solve the problem that the dermoscopy image is difficult to collect and the color display is abnormal,this paper gives the dermoscopy image batch acquisition and image color correction method.The dermoscopy image batch acquisition method utilizes the dermoscopy image dataset interface provided by the International Skin Imaging Collaboration to design a data crawling method,realizes batch acquisition of the dermoscopy image,and uses the Shades of Grey algorithm to perform color constancy correction on the acquired image,this correction method can restore the dermoscopy image with abnormal color to the image under natural white light.3.This paper presents an end-to-end dermoscopy image segmentation network to achieve semantic segmentation of dermoscopy image.The network is divided into a down-sampling part and an up-sampling part.The down-sampling part extracts the feature map of the dermoscopy image by using the densely connected convolution,and the up-sampling part uses the deconvolution to restore the feature maps of different sizes to the size of the input image,and realizes the image segmentation.The segmentation network initializes the down-sampling parameters by using the transfer learning method,and uses the fine tuning technique to train the parameters of the deconvolution kernel,thereby improving the efficiency of network training.This paper also designs a loss function based on the dice coefficient for the network,making it more suitable for the dermoscopy image segmentation task.4.Based on the image segmentation of dermoscopy image,a classification algorithm of dermoscopy image based on ensembled convolutional neural network is presented.Firstly,according to the segmentation result,we removes the background part of the dermoscopy image,retains only the lesion area image,and then uses the lesion area image to train multiple convolutional neural networks to extract the lesion feature vector,and uses the ensemble learning method to classify the feature vector to achieve the purpose of dermoscopy image classification.The method also uses the pre-trained parameters on the natural picture to initialize the convolutional layer of the convolutional neural network according to the transfer learning strategy,so that the convolutional neural network can be effectively trained in the case of limited training data to avoid over-fitting of the model.In order to verify the effectiveness of the dermoscopy image segmentation and classification method,a number of comparative experiments were performed using the dermoscopy image dataset published by ISIC.In the dermoscopy image segmentation experiment,the dermoscopy image segmentation network given in this paper is compared with many classical image semantic segmentation methods.The Jaccard index between the prediction segmentation result of the network and the real segmentation result reaches 83.8% which is improved by more than 6% compare to other method.In the dermoscopy image classification experiment,we verify the effect of color constancy correction method,image background removal processing and the ensemble convolutional neural network algorithm through multiple sets of contrast experiments.The experimental results show that the color constancy correction and background removal processing of dermoscopy image can obtain better classification results,and the classification accuracy rate of the dermoscopy image classification algorithm based on the ensembled convolutional neural network reaches 89.2%.which achieved the state-of-the-art performance on ISIC2018 datasetThe results of multiple experiments show that the method presented in this paper can more accurately and efficiently perform dermoscopy image segmentation and classification,and has certain reference value for clinical medical diagnosis and medical research.