Machine Learning Based Skin Disease Image Classification

Recent years,computer-assisted medical image processing and analysis techniques have been widely used.With the development of machine learning technology,the study of medical image processing technology based on machine learning has become a hot issue in the field of computer-aided diagnosis.Skin cancer is currently a very common type of cancer,with patients in all age groups.The diagnosis process of skin cancer is complicated,which requires the doctor to observe and judge first,and then perform biopsies detection under the microscope.Therefore,a more accurate image classification algorithm for skin diseases will provide great help in the diagnosis of skin cancerThe automatic classification of skin image features has many difficulties,but deep convolutional neural networks provide a technical basis for this process.This topic aims to study the classification and prediction methods of dermatological clinical images based on deep learning and compare it with traditional machine learning methods.Build an independent skin disease database to train and optimize neural networks to improve the accuracy of skin disease classification.The main innovations and contributions are summarized below:1.Given that the large-scale clinical datasets of dermatological diseases have not been published,a dermatological clinical image dataset containing a total of 4702 images of 90 common skin diseases has been established in this paper.The data set is highly diverse:species diversity,appearance diversity,and attribute diversity.At the same time,we performed preprocessing operations such as image denoising,image enhancement,and hair removal on the data set,making the data set have excellent classification performance.2.Aiming at the poor classification effect of single feature extraction method in traditional classification methods of machine learning,by analyzing the image features of clinical dermatoses images,this paper proposes a multi-feature fusion method that uses a Canonical Correlation Analysis(CCA)method to fuse single features.After training the SVM classifier,the classifier is optimized by selecting the kernel function with better performance.Experiments show that the multi-feature fusion method improves the classification accuracy.3.By deep learning neural network and using transfer learning,the effects of different network models and different training methods of transfer learning on the classification results of different types of skin diseases were compared.At the same time,a new type of multi-scale neural network model is proposed,which uses features of different scales for feature extraction and neural network classification.Experiments show that the multi-scale neural network model improves the classification accuracy of different types of dermatological images.