Research On Boundary Guided Semi-supervised Segmentation Method Of Breast Tumors

Breast cancer has become the most common cancer in the world,and the mortality can be greatly reduced if detected and treated early.As the incidence and number of breast cancers continue to rise,the inability to perform screening and early diagnosis manually has resulted in many potential cases not being detected at an early stage and missing the best time for treatment.Computer-Aided Diagnosis(CAD)systems can help physicians make rapid diagnoses and conduct extensive breast cancer screening.However,the diagnostic performance of CAD systems for breast cancer is dependent on the precise Region of Interest(ROI),precise segmentation of breast tumors ROI becomes a critical factor limiting its performance.The performance of fully supervised breast tumor segmentation methods based on deep learning are limited in the field of breast tumor segmentation due to issues such as variable target size,fuzzy boundaries,data imbalance,and scarcity of labeled data.This paper proposes a boundary-guided semi-supervised breast tumor segmentation network based on a self-built breast MRI tumor segmentation dataset to address these issues.The main contributions are as follows:(1)This paper builds a boundary-guided fully supervised breast tumor segmentation network called BGSegNet based on boundary to solve the problems of variable tumor morphology,blurry boundary,and data imbalance.To fully utilize tumor boundary information and alleviate the problem of data imbalance,the network employs the classical encoder-decoder structure and adds dynamic boundary loss to the cross-entropy loss.Furthermore,an atrous spatial pyramid pooling module is added to the encoder’s final layer to obtain richer multiscale features,and a composite attention module is embedded between each layer of the decoder to achieve both meaningful feature enhancement and noise suppression.Finally,the encoder’s features are directly passed to the decoder in order to achieve the fusion of detailed and abstract features.On the independent test set,BGSegNet’s Dice coefficient,Sensitivity,Specificity,and Hausdorff distance are 0.748,0.886,0.998 and 3.124,respectively.In terms of Dice coefficient and Hausdorff distance,it outperforms the UNet,FCN8 s and AUNet.Experiments demonstrate the efficacy of the proposed boundary-guided network for breast tumor segmentation,laying the foundation for the automated CAD system for breast tumors.(2)This paper builds a semi-supervised breast tumor segmentation network based on the idea of generative adversarial networks to solve the current problem of lacking labeled data.The network is consists of two relatively independent networks,the BGSegNet network in the first stage and the discriminator network in the second stage.Self-learning of unlabeled data can be accomplished by including a lightweight fully convolutional neural network in the training stage without changing the structure of the segmentation network.The proposed semi-supervised network’s Dice coefficient,Sensitivity,Specificity,and Hausdorff distance on the independent test set are 0.819,0.900,0.998 and 2.850,respectively,better than fully supervised networks like UNet and AUNet using the same number of labeled data,and also better than the comparison semi-supervised networks.Experiments show that the semisupervised network proposed in this paper can not only fully utilize the supervised information in the labeled data,but also extract information from the unlabeled data,both of which interactively boost the performance of segmentation network.In summary,this paper fully utilizes the information in the target boundary and unlabeled data to improve the accuracy of breast tumor segmentation,which is critical for realizing the automated mass screening of breast tumors.