Skin Lesion Image Segmentation Algorithm Based On Deep Learning

Skin lesion image segmentation is a field of medical image segmentation,and it is important for dermatologists to diagnose skin diseases.In recent years,with the upsurge of artificial intelligence,research on skin lesion image segmentation by means of deep learning has broad application prospects.For the shape and texture of the lesion area in the skin lesion image,and the relatively low contrast of the surrounding lesion area,the segmentation result is greatly deviated.This paper proposes three segmentation algorithms for skin lesion image segmentation from different perspectives based on the deep learning method.First,this paper proposes a skin lesion image segmentation algorithm based on improved fully Convolution Network.The algorithm is based on the Fully Convolution Network(FCN),and introduces the Jaccard-Diceloss loss function instead of the softmax loss function in the FCN to obtain a new partition network,namely the improved fully convolution network(IFCN).The IFCN will not only inherits the advantages of the FCN,but also solve the problem that the lesion area(ie,foreground)of the lesion image is too different from the non-lesion area(ie,background)during the training prediction process.Through the training and prediction of IFCN,the evaluation index values and the segmentation results of the skin lesion images are best.It shows the effectiveness and significant effect of the IFCN on the image segmentation of skin lesions.Secondly,This paper proposes a skin lesion image segmentation algorithm based on a multi-scale fully convolution dense block network(MSFCDN).The network uses a multi-scale structure,a full convolution structure,and a dense block structure.It solves the problem of insufficient image feature extraction.Through the experiment,the evaluation index values and the lesion segmentation area maps are all expected.And the segmentation effect of the lesion area is fine and accurate.It indicates the validity of the MSFCDN.Finally,This paper proposes a skin lesion image segmentation algorithm based on multi-scale encoder-decoder network(MSEDN).The network adopts a multi-scale input and coder-encoder structure.At the same time,output a bilinearly interpolated intermediate prediction layer to the decoding network for cascading input in the pool2 layer of the coding network.The MSEDN is used to solve the problem that the general network training speed is too slow,the model parameters are stored too large,and the feature extraction is insufficient.Through the experiment,the evaluation index values and the lesion segmentation area maps are all expected.It shows that the MSEDN has a good accuracy for segmentation of skin lesion images.