Research And Implementation Of Skin Cancer Auxiliary Diagnosis Method Based On Convolutional Neural Network

Skin cancer is a common cancer that occurs not only on the surface of the skin but also on the skin appendages,with a high mortality rate.Skin cancer is mainly divided into melanoma and non-melanoma,among which melanoma is the most aggressive and fatal skin cancer.The 5-year survival rate of patients diagnosed with advanced melanoma is only 17%,but if detected and treated as early as possible,the 5-year ultimate cure rate is over 98%.Therefore,it is critical to detect and treat skin cancer in time.With the rapid development of modern medicine and computer technology,Computer-Aided Diagnosis technology has become an indispensable tool in the field of medical imaging,which can assist doctors in image analysis and diagnosis.It is of great significance to improve the accuracy of early diagnosis of diseases.Computer-aided diagnosis technology based on deep learning has been widely used in skin cancer image classification,but it still has the following shortcomings: the model suffers from overfitting problems in the case of unbalanced data samples or data scarcity;some feature interactions between layers of different convolutional layers are easily ignored and redundant information exists in the features extracted by the model.To address the above problems,this paper designs a classification method that can efficiently detect skin cancer based on Convolutional Neural Network.The method is based on the Efficient Net V2 model,and proposes an improved Efficient Net V2 model based on feature fusion and random forest and builds a skin cancer detection system,which is tested to prove that the system can provide convenience for doctors to diagnose skin cancer.The main research contents of this paper are as follows:1.Aiming at the problem that the classification model ignores some feature interactions between layers,resulting in insufficient feature utilization,an improved network architecture based on the Efficient Net V2 model is proposed.By modifying the model structure,the architecture removes the last few layers of the model and retains the feature extraction part of the original model.Then,after the improved model,hierarchical bilinear pooling is added for feature fusion.By fusing features of different levels,the interaction relationship between some features between layers can be captured,and the expressive ability of features can be enhanced,thereby improving the accuracy of model classification.2.In order to solve the problem of overfitting Convolutional Neural Network due to insufficient sample size and unbalanced dataset classes,the backbone of the proposed model is trained using a migration learning approach and the random forest algorithm is used for classification.The backbone network of the algorithm uses the method of pre-training to obtain the initial weight,and then the initial weights are applied to the HAM10000 dataset using model fine-tuning.This approach can effectively utilize the knowledge of the pre-trained model and reduce the training time and data volume.In contrast,random forest can improve the classification accuracy of the model by integrating the results of multiple decision trees.When constructing decision trees,random forest uses random feature selection and random sample selection to increase the diversity of each decision tree to avoid overfitting,and it can also balance the dataset and improve the generalization ability of the model.3.To address the problem of redundant information in the features extracted by the Efficient Net V2 model,this paper adds the efficient channel attention mechanism to feature selection or weighting before hierarchical bilinear pooling to retain features that are more important for the classification task and reduce the interference of redundant and noisy information.Meanwhile,the efficient channel attention module can achieve the purpose of enhancing the model perceptual field by increasing the nonlinear transformation of weights without increasing the number of parameters of the model,thus improving the model performance without affecting the model training speed.4.According to the classification model proposed in this paper,a skin cancer detection system was designed and implemented.The system includes user login module and lesion detection module.The test results show that the system has good usability and reliability,which can effectively help doctors in the diagnosis of skin cancer and provide better medical services to patients.