Research On Dermoscopy Image Lesion Detection Based On Machine Learning

With the improvement of people quality of life,health issues are receiving more and more attention.As a common disease,skin disease has a great impact on the lives of patients.Without timely and effective diagnosis and treatment,mild patients' skin will be damaged,severe patients will develop stubborn skin diseases or even cancer,the detection of skin diseases mainly relies on artificial naked eye observation,resulting in the diagnosis rate being greatly affected by the subjective factors of professional doctors.How to use artificial intelligence to improve the accuracy of diagnosis is an important problem faced by experts and scholars.This paper focuses on the dermoscopy image lesion detection algorithm based on machine learning and conducts experimental verification.The specific research work is as follows:Firstly,a HSV(Hue,Saturation,Value,HSV)color space and new novel full convolution dermoscopy image lesion region segmentation method is proposed.The fusion of HSV color space and Gaussian weighted mean segmentation strategy to improve the stability and anti-interference ability of the segmentation algorithm of dermoscopy image lesion area,and the introduction of GrabCut segmentation method to further improve the segmentation effect;at the same time,the traditional U-Net feature extraction structure is transformed into Lightweight feature extraction structure,to explore a lightweight semantic segmentation network ML-Unet with mixed modal input,through the introduction of convolution transform module to solve the problem of potential semantic differences in skip connection in U-Net structure,and to encode The output features of different scales in the decoder are bilinear difference,which provides the decoder with richer semantic features and improves the segmentation accuracy of the model.Secondly,the dermoscopy image lesion recognition method based on feature descriptors and novel lightweight convolutional neural network is studied.Based on mainstream feature descriptors SIFT(Scale Invariant Feature Transform,SIFT),SURF(Speeded-Up Robust Features,SURF),LBP(Local Binary Patterns,LBP)and local depth feature,a classifier based on visual dictionary and feature histogram training are constructed to improve the recognition rate of dermoscopy image lesions;a feature extraction network is constructed based on a lightweight convolution module,At the same time,the principle of fine-grained classification is introduced to measure the similarity of the feature vectors output by the feature extraction network to improve the model feature distinguishability and recognition accuracy,and based on the Grad-CAM++algorithm and T-SNE algorithm,the proposed disease recognition model is visualized and analyzed from the perspectives of feature activation map and feature clustering.Finally,a dermoscopy image lesion detection system based on the PyQt platform is designed.The QtGui module was used to construct the display interface of the detection system,and the QtWidgets module was used to develop the function control keys of the detection system,and the actual collection of skin disease images was used to verify the feasibility and effectiveness of the system.