Research Of Automatic Segmentation And Recognition Of Skin Lesion Images Based On Deep Convolutional Networks

Melanoma,as a representative of malignant skin pigmented lesions,have an extremely high mortality rate,and in early stage,it is easily confused with nevus and other high incidence of benign skin pigmented lesions,which resulting in delayed treatment.At present,diagnosing skin pigmented lesions by dermatoscopy imaging relay on the doctor’s clinical experience.And it’s time-consuming and may be subjective.The computer-aided diagnosis system(CAD)for skin pigmented lesions can alleviate the necessitous medical resources,relief the workload of doctors,and improve the efficiency of diagnosis.Consequently,it is benefit for patients to strive for valuable treatment opportunities.However,the clinical manifestations of same skin pigmentation lesions are diverse largely but different lesions may be appearing highly similar.Usually,The boundary is blurred.These bring significant challenges for computer diagnostic systems realizing of skin pigrmented lesions.This paper studies the computer-aided diagnosis of skin pigmented lesions in two aspects:skin pigmentation lesion segmentation and recognition based on convolutional neural networks.The main research work of this paper is as follows:(1)As a modified UNet,This paper proposes a two-stage convolution neural networks based on multi-level feature fusion.By fusing multi-level feature of the encoding part can promote the recovery of the spatial position information in the decoding part processing by introducing spatial position information feature into the header of the decoding network part early.The multi-level feature fusion of the decoding part provides a skip connection from the classifier to the deep layer of the network,and accelerates the flow of gradient information,thereby improving the training of the model.Aiming at the problem of uneven gradation distribution of the lesions area,deformable convolution is added in the encoding part to adaptively enlarge the receptive fields.Finally,by the two-stage method,the segmentation effect of the smaller lesion area is further improved.By these ideas,the baseline segmentation model based on ResNet50 gets a 1.3%improvement in jaccard index on PH2 Dataset.(2)This paper designed an automatic recognition model of skin lesions based on convolutional neural networks.The multi-source image data is preprocessed by a color consistency algorithm.Aiming at the problem of unbalanced class distribution due to the different incidence of skin lesions,the smoothed logarithmic class frequency weight is used to adjust the weight of category in the loss function.Finally,we study the hard example learning problem in the classification task.The Focal loss function and online hard example mining(OHEM)method,which are proposed in the detection task,are used to enhance the learning of hard samples in the training process,thus improving the performance of the model.It achieves a 0.9688 F1-Score in Melanoma Dataset.