Research On Melanoma And Pediatric Pneumonia Recognition Based On Deep Learning

Melanoma is a very harmful skin disease.Its treatment effect is good in the early stage,but the late treatment effect and prognosis are poor.Therefore,the early diagnosis of melanoma is very important to reduce its mortality.In addition,pneumonia is the leading cause of death in children all over the world,so timely diagnosis of pneumonia is particularly important to improve the survival rate of patients.Dermoscopic imaging and chest X-ray scanning are common methods used for skin diseases and pneumonia examinations.These medical imaging techniques can help doctors better observe the diseased area.Clinically,the diagnosis of diseases is basically performed by manual reading.However,due to the subjective influence of doctors’ personal experience and level differences and the uneven distribution of medical resources in different regions,the accuracy of diagnosis cannot be guaranteed.Compared with the traditional way of manual image reading,automatic diagnosis technology based on medical images can effectively solve the above mentioned problems and assist doctors in diagnosis.Traditional machine learning-based medical image processing algorithms are complex and require relevant professional knowledge.In contrast,deep learning algorithms can end-to-end learn and directly generate results on input data.Based on the above analysis,according to the characteristics of dermoscopic images and pediatric chest X-ray images,this thesis designs different convolutional neural network models to identify melanoma and pediatric pneumonia.The specific research has the following aspects:(1)In the field of medical image processing,the amount of data that can be used for network training is not sufficient,and the scale of the data greatly limits the performance of the network.In order to alleviate the impact of insufficient data,this thesis adopts the method of transfer learning using the pre-trained VGG19 network as the basic model.In addition,this thesis improves the VGG19 network,and adopts a global average pooling layer after the convolutional layer to form the VGG19-GAP basic network.The parameter amount of VGG19-GAP network is about 14.4% of VGG19 network.And in the task of melanoma recognition,the AUC value obtained by the VGG19-GAP network is 1.4% higher than that of the VGG19 network.(2)For melanoma recognition,the difference between the skin lesion categories is small and the normal convolutional neural network is not sensitive to the lesion area,resulting in poor recognition performance.Based on this,this thesis proposes an attention-based network(AF-CNN)for melanoma recognition.On the basis of the VGG19-GAP network,attention blocks are added to enable the network to adaptively focus on the lesion area in the dermoscopic image and reduce the influence of the background areas.In addition,in order to increase the discriminative ability and richness of the extracted features,this thesis fuses the attention features of the high and low layers of the network so that the network can make a more accurate classification based on the extracted features.And the proposed AF-CNN model is evaluated on the ISIC2017 dataset.Compared with the two basic network models of VGG19 and VGG19-GAP,the AUC value of AF-CNN has increased by 4.5% and 3%,respectively.In addition,compared with the first place in the ISIC2017 competition,the AUC value has increased by about 1.3%.(3)Compared to dermoscopic images,chest X-ray images are more complicated,and various organs overlap.The chest X-ray image covers a wide area,and only a part of the entire image is the area related to the diagnosis of pneumonia,the recognition performance of the network will be affected by irrelevant regions.Therefore,this thesis proposes a two-branch GL-CNN network based on feature fusion.The first branch uses the entire pediatric chest X-ray image as input,and adopts weakly supervised localization method to generate the localization map in the process of forward propagation.Then,the localization map generated by branch one is used to find the discriminative region in the entire image.This area is cropped and enlarged and used as the input of branch two to extract more detailed features.Finally,the features of the entire image and the detailed features of the local area are fused to strengthen the discriminative ability of features.The proposed network is evaluated on the pediatric chest X-ray image dataset.The recognition accuracy and AUC value obtained by the network are 0.9279 and 0.9827,respectively.Compared with AF-CNN and other methods,the recognition performance of the network has been improved.Different from natural images,medical images are more complex,with small differences between different disease categories and insufficient data,which makes classification more difficult.For melanoma recognition and pediatric pneumonia recognition,this thesis proposes corresponding improved models.Compared with other methods,the performance of the proposed models have been improved.It has a good reference value for the study of melanoma and pediatric pneumonia recognition technology,and also has important application value.