Research On Image Segmentation Method Of Breast Mass Based On Spatial Adaptive Adversarial Network

With the changes in eating habits and the development of medical diagnosis,the number of breast cancer patients is increasing year by year.Breast cancer,as the most common cancer among women,has become a major public health problem in society,and the second leading cause of female death.Early detection,early diagnosis and early treatment are the key to improving the cure rate and reducing the mortality rate.Breast tumor is one of the most common pathological features of breast cancer in clinical practice.They have different sizes and shapes,complex edges,and diverse features.Mammogram is an important basis for identifying breast tumors.In clinical practice,most radiologists provide reliable opinions to assist diagnosis based on X-ray photos,but manual screening is time-consuming and laborious,and is limited by different medical standards and other subjective factors.Therefore,how to automatically segment breast tumors in mammograms is very necessary and critical in the early diagnosis of breast cancer.In recent years,deep learning is in the ascendant.In medical image processing,especially in the field of image segmentation,methods based on deep learning have achieved remarkable results,and their segmentation accuracy has surpassed traditional segmentation methods.Various network models such as FCN,U-Net,Deep Lab,GatedSCNN,etc.have emerged one after another,which has injected strong momentum into the development of medical image segmentation.In this article,we proposed a breast mass segmentation model based on a spatial adaptive adversarial network to automatically segment breast mass regions in mammograms.In mammograms,breast tumors are usually covered by the surrounding dense tissues,with irregular shapes,uneven edges,and low contrast,which posed a greater challenge for automatically segmenting accurate tumors.In order to obtain more refined and accurate segmentation results,in the overall architecture,we use the model framework of generative adversarial networks which contains a segmentation network and a discrimination network.The two networks are used for adversarial training,and the results of the discrimination network are used to supervise the segmentation network learning direction.and then achieve the goal of optimizing segmentation results.In addition,the segmentation task is a time-consuming mission,and adversarial training takes a long time to converge.Because the traditional convolution operation would drop the key semantic information,we added a spatial adaptive normalization layer to the discriminant network to capture and maintain the semantic information in the corresponding segmentation mask,and then guide the learning direction of the discriminant network.In addition,we employed a hybrid loss function composed of adversarial loss,segmentation loss,and perceptual loss.Because in the breast mass segmentation task,the masses in the image are relatively small and the background area is relatively large,so in order to improve the category imbalance problem,we utilized Dice loss as the segmentation loss function,instead of the classic L1 loss.At the same time,in order to ensure that the generated result is consistent with the real label in terms of semantic features,we also introduced the perceptual loss to measures the gap between the feature maps.In experiment,the combination of segmentation and generation of the confrontation network requires a large number of convolution operations,and the amount of parameters is huge.In order to reduce model parameters,we introduce deep separable convolution into the U-Net network in the experiment.We conducted experiments on two most commonly used public mammography mass segmentation datasets(INBreast and CBIS-DDSM).The experimental results show that our method performs better than other latest models.Its Dice index reached 81.15% in the INbreast data set and 83.35% in CBIS-DDSM,which fully proved the effectiveness of the model.