Research On Dermoscopic Images Segmentation Method Based On Attention Mechanism

Skin is the largest organ of the human body,and there are many kinds of pathological changes.Pigmented skin disease is one of the most common skin diseases,mostly caused by changes in pigments.Melanoma is caused by malignant lesions of melanocytes,and is the most serious malignant tumor in pigmented skin diseases.Its early symptoms are easily confused with other benign skin tumors and are not easy to detect,which causes patients to miss the best time for diagnosis and treatment.Even for professional dermatologists,it is difficult to diagnose melanoma directly with the naked eye.At present,dermatologists usually use dermoscope to make clinical diagnosis of patients.Although dermoscope can magnify the skin lesion area and improve the identification of the lesion area,this diagnosis method still highly depends on the professional ability of the dermatologist,which is time-consuming and laborious.The computer-aided diagnosis system can improve the efficiency of detecting skin diseases,and the accurate segmentation of the lesion area of the dermoscopic image can improve the accuracy of the diagnosis system to identify skin diseases,so as to help doctors make better diagnostic decisions and gain more treatment time for patients.The main contents of this thesis are as follows:(1)A dermoscopic image segmentation model based on serial attention is proposed.In order to make full use of the information aggregation ability of attention mechanism,spatial attention and channel attention are combined in different forms in this thesis,and the combined modules are embedded into the U-shaped network.The U-shaped network uses the Dense Net121 as the encoder,and uses the Dense Atrous Spatial Pyramid Pooling module to obtain densely multi-scale information,which improves the model’s ability to capture global and local features.At the same time,for the problem of inconsistent illumination in ISIC2017 dataset,this thesis uses the gray world algorithm to perform color balance processing on the dermoscopic image to reduce the impact of the lighting environment on the color display.The experimental results show that the serial combination of channel attention followed by spatial attention has better segmentation effect than other combinations.The Jaccard Index value of 0.7692 was obtained on the ISIC2017 dermoscopic image dataset.(2)A dermoscopic image segmentation model based on attention correction is proposed.In this thesis,the dense block is used as the basic component of encoder and decoder of the U-shaped network.At the same time,the prior knowledge obtained by the pre-training network combined with the serial attention module is used to correct the features extracted from the encoder,which improves the segmentation performance of the model.In order to make the model give enough attention to the boundary,a composite loss function combining the regional loss and the boundary distance penalty is designed.In the process of gradient descent,the prediction of the lesion region outside the boundary will obtain a positive gradient,which will reduce the prediction probability of this area.On the contrary,the prediction of the lesion region inside the boundary will get a negative gradient,thus increasing the prediction probability of this region.The proposed model and loss function were applied to the ISIC2017 dermoscopic image dataset,and the Jaccard Index value of 0.7836 is obtained,which is a 1.86% improvement in Jaccard Index value compared with the ISIC2017 challenge champion model.