Research On Melanoma Classification Algorithm Based On Deep Learning

Melanoma is a highly malignant skin tumour,and early detection and treatment is very important to the prognosis of patients.However,current diagnostic methods mainly rely on dermatologists’visual observation and histopathological biopsy,the former of which is subjective and the latter of which can cause unnecessary pain to patients.Thanks to the development of deep learning technology,the use of computer-aided diagnostic systems can help physicians analyze and determine skin lesions more quickly and accurately,thereby increasing their efficiency and reducing their workload.Aiming at the problems of a small amount of training data of skin lesion samples,inter-class similarity,wide sources of data samples,and interference of irrelevant background information,the research in this thesis focuses on the recognition and classification of skin lesions in dermoscopic images and images captured by mobile devices,and the specific work includes the following three points:（1）Research on melanoma classification method based on improved ResNet50 network:Because of the time-consuming and laborious problem of using dermoscopy to examine melanoma in clinical practice,this thesis proposes a melanoma classification model based on migration learning and an improved ResNet50 model.Specifically,firstly,data enhancement is performed on the classification dataset to reduce the effects of uneven distribution of samples in the dataset and hair obscuration information.Second,to improve the classification performance of the model,the ResNet50 is improved,i.e.,the input backbone of the network is redesigned to reduce the number of parameters of the model while keeping the perceptual field unchanged;the residual blocks are optimized to enhance the robustness and expressiveness of the model;and the CA attention mechanism is added to the network to make the model pay more attention to the important features and thus improve the classification accuracy.Finally,the model is trained with transfer learning to improve the performance of the model on the classification task.The experimental results show that the accuracy and₁-score of the proposed method increase to 95.4%and 95.6%,respectively,after migration learning.（2）Two-stage classification of melanoma based on a lightweight segmentation network:Since the existing algorithms are directly classifying dermoscopic images and are not suitable for automatic segmentation and classification of skin lesions on mobile devices,this thesis proposes a two-stage classification method for melanoma based on a lightweight segmentation network.Specifically,for the segmentation part,this thesis introduces a multi-branch fusion technique to improve the U-Net and reconstructs the U-Net network using a re-parameterisable convolution module to achieve fast segmentation of skin lesion boundaries.For the classification part,this thesis first embeds the Sim Am attention module into the Middle Flow of Xception,and then uses a multilayer perceptron with dropout added instead of the fully connected layer in Exit Flow,while the Focal Loss loss function is used to train the network to alleviate the imbalance between hard and easy samples learning.Further,a hybrid data enhancement strategy including Cut Mix is used for the classification dataset to improve the generalization ability of the model.The experimental results show that the Io U value and Dice coefficient of the proposed method reach 80.03%and 87.74%respectively on the ISIC2018 segmentation dataset with the lowest model complexity,and the accuracy and₁-score of the proposed method increase to 95.61%and 95.76%respectively on the HAM10000 dataset.（3）Based on the above algorithm,this thesis establishes a skin lesion recognition and diagnosis system to achieve intelligent diagnosis of skin lesion images from different sources.The system includes multiple modules,including image upload storage and preprocessing module,lesion area segmentation module,lesion recognition and diagnosis module,result display module,and skin lesion science popularization module,which can achieve more accurate and timely diagnosis and treatment of skin lesions.