Research Of Medical Skin Image Classification Algorithm Based On Deep Learning

Malignant skin tumours pose a great threat to patients’ health.Due to the limitations of existing diagnostic techniques such as poor accuracy and invasive operations,as well as the high similarity between malignant skin lesions and other skin lesions,the clinical diagnosis of malignant skin tumours has low precision,high misdiagnosis rate and low efficiency.The use of computer algorithms for automatic classification of skin medical images can effectively improve the efficiency of clinical diagnosis.(1)The quality and quantity of image data used for training the model directly affects the performance of the model,and there is still room to improve the accuracy of the model due to insufficient data and category imbalance.(2)Most of the current models are based on dermatoscopic images of the affected area,while there are few studies and datasets of clinical images of the skin for diagnosis.Compared to dermoscopic images,the background in skin clinical images is more complex,the lesion area is smaller,and the ambient light and noise in the non-lesion area are more challenging.To address the first problem,this paper proposes a two-stage framework called G-DMN,which uses the CycleGAN extended dataset and Dense-MobileNetV2(DMN)to achieve automatic classification of skin lesion images.In the first stage,this paper uses CycleGAN for data expansion and proposes a new image pairing strategy: learning the transformation of majority class images to minority class images by CycleGAN,and then generating minority class images to balance the dataset.In the second stage,a lightweight model called DMN is proposed in this paper.By improving MobileNetV2,it enhances feature reuse by increasing the width of the network and allows the network to focus on focal regions at different scales.The original training set is combined with the generated images and used to train the DMN.experiments demonstrate that the proposed method is lighter and more effective than classical classification methods,achieving significant performance improvements.The article addresses the second problem by constructing a new Clinical Skin Lesion Images(CSLI)dataset and proposing a classification model based on a Double Branch Net(DBN),which contains two branches,the original network and the fusion network,and this paper proposes a CFEBlock to extract the The proposed CFEBlock extracts the common features between the neighbouring levels of the original network branches,and then fuses them with the fusion network branches through the proposed Fusion Block,and finally obtains the prediction results by weighted summation.Finally,the effectiveness of each DBN module is compared with other deep convolutional neural networks on the CSLI dataset.The metrics of the model on different categories of diseases were analysed,ultimately showing that the network performed better overall.