Semantic Feature Guided Regression Model For Breast Cancer Cellularity

Breast cancer(BCa)has become one of the dominant causes of death from cancer for women around the world in recent years.For the reason that the tissues in the breast are relatively loose and there are a large number of dense lymph nodes around the breast,the tumor cells are easily transferred to other organs of the body along with the lymph and the blood,leading to the distant metastasis of the BCa.The patient with distant metastasis of BCa should be treated with the Neoadjuvant Therapy(NAT)until the Residual Cancer Burden(RCB)is reduced to a certain level,which means that the patient has the operation indication,and then the tumor resection can be performed.The calculation of the RCB requires six parameters,and the most complicated and difficult one is the tumor cellularity(TC),which is pathologically defined as the ratio of the nucleus area of the tumor cell in the tumor bed(TB)to the pathologic slicing image observed by the microscope.Manual calculation of the TC is time-consuming,error-prone and imprecise due to the small size and dense distribution of the nucleus.Hence,designing algorithms to estimate TC automatically can reduce the burden of pathologists and improve the calculation efficiency of TC and RCB.At present,automatic algorithms for TC estimation can be divided into two categories:two-stage TC estimation method based on nucleus segmentation,and one-stage TC estimation method.The former is on the basis of the pathological definition of the TC,calculating the TC according to the segmentation result.The latter utilizes the deep learning model to predict TC directly.The former is more similar to the pathologist’s workflow.Thus,the former has better interpretability.However,the performance of the former depends largely on the performance of the segmentation model.For purpose of improving the performance of the algorithms for TC estimation,this thesis takes advantage of the popular machine learning algorithms and proposes a semantic feature guided regression model for TC estimation.Firstly,this thesis modifies the U-Net to implement the cell nucleus segmentation network based on weakly supervised learning,which can segment the nucleus accurately utilizing datasets with annotations of the position coordinates and the category of the nucleus merely.Specifically:First,two types of pixel-level segmentation annotations,namely contour annotations and position annotations,are generated according to the weak annotations.Then,the deeply supervised mechanism is introduced into the modified U-Net,and the loss of the segmentation network is calculated based on two types of generated annotations.To evaluate the performance of the segmentation network accurately,under the guidance of the pathologists in Zhongnan Hospital of Wuhan University,we label the pixel-level annotations for the test dataset manually,and utilize the annotations as the ground truth to evaluate the performance of the proposed segmentation network.The experimental results demonstrate that the number of false positive examples of the foreground cell nucleus is significantly reduced and the mean Intersection Over Union(m IOU)of the segmentation network is increased from 54.42%to 58.80%after adding position annotations to assist contour annotations to train the segmentation network.Then,this thesis proposes a semantic feature guided regression model for TC estimation on the basis of the cell nucleus segmentation network and the TC regression network.Given the fact that the segmentation network can extract the high-level semantic feature of the cell nucleus which is instructive and meaningful to TC estimation,the multi-scale feature maps from the decoder part of the segmentation network are fused with the feature maps from the corresponding layers of the TC regression network.Feature maps from two networks are concatenated in the channel dimension firstly,and then convolution operation is conducted to reduce the number of channels of the feature map.Furthermore,this thesis adopts the theory of the label distribution learning(LDL)and designs the TC regression network based on the deep expectation estimation(DEE).Considering that the values in the neighborhood of the TC label value can describe the image reasonably and correctly to a certain extent,non-zero weights are assigned to the correct labels at corresponding positions with the pseudo-gaussian distribution.The floating-point TC value is encoded as a 101-dimensional label distribution vector to avoid over-fitting on the current labeled dataset.The LDL method and weighted multi-class cross entropy are utilized in the process of training the regression network.The result of the DEE is adopted as the TC estimation value.After combining semantic feature guidance with DEE,the performance of the model is improved,which proves the validity of the model.When we utilize the Res Net101 as the backbone,the prediction probability(p_k)of the TC estimation model is 0.9335 whereas the p_k is 0.9154 when modeling the task as a regression problem directly.In conclusion,a semantic feature guided regression model for breast cancer cellularity is proposed in this thesis on the basis of the modified U-Net and the theory of the LDL.The model consists of the cell nucleus segmentation network,feature map fusion module and TC regression network.The training of segmentation network only needs weak labeling,and therefore it has good generalization.The model implements the automatic estimation of TC,reduces the workload of pathologists,and improves the calculation accuracy and efficiency of TC and RCB.What’s more,research on the algorithms for TC estimation is of great significance to the design of the treatment for BCa patients,the clinical promotion of RCB index and the development of BCa chemotherapy drugs.