Research On Automatic Diagnosis Of Breast Cancer Digital Pathological Images Based On Deep Convolutional Networks

Today,breast cancer is the cancer with the highest incidence and mortality rate worldwide,which seriously endangers people’s physical and mental health.As a heterogeneous disease,histopathology is the "golden standard" for diagnosing breast cancer.Pathologists usually judge the stage or grade of cancer by observing the biopsy of tissue under microscope.The structure of breast tissue is extremely complicated,so the manual diagnosis is time-consuming and labor-intensive,and sometimes there would be missed diagnosis and misdiagnosis.At the same time,the histological diagnosis is affected by the differences of personal experience and judgment criteria,and the diagnostic consistency rate is low.Therefore,it is of great value and practical significance to develop a breast cancer automatic diagnosis system to assist the pathologist to improve the accuracy and efficiency of diagnosis.With the development of deep learning,deep convolution networks have gradually become an important tool to recognize digital pathological images.In this paper,the deep convolution networks are utilized to carry out the automatic diagnosis of hematoxylin& eosin-stained breast histopathological images on the two tasks of cancer area detection and histological grading.The main contents are as follows:(1)A semantic segmentation method based on lightweight dual-path network is proposed to divide whole-slide images of breast cancer into four regions: normal,benign,in situ carcinoma and invasive carcinoma.According to the characteristics of breast cancer whole-slide images,such as high resolution,high intra-class difference and low inter-class variance,the dual-path network adopts a light-weight framework structure combined with dilated convolution and channel attention mechanism to extract,filter and combine discriminative information at different levels.The module achieves 68.12% m Io U and 87.02% pixel accuracy on ICIAR2018 BACH dataset,performing excellent in accuracy as well as efficiency.(2)An integrated system based on multiple classifiers is proposed to automatically assess the atypia of breast cancer.The residual convolution network Resnet50 is used as the base classifier under single resolution,and then the classification results under each resolution are combined strategically by metadecision tree.Transfer learning is adopted when training base classifiers to solve the limitations of small-scale samples for deep network learning.In addition,color standardization is used to preprocess the images collected by different devices to improve the generalization ability of the model.The model achieves 88.00% accuracy and 86.45% sensitivity on ATPIA14 dataset,which is more competitive than other latest algorithms.(3)A method of nuclear segmentation based on hybrid-attention nested U-Net network is proposed to optimize the automatic scoring system for breast cancer,in order to reduce the interference of cytoplasm and other regions in breast tissue.The combined structure of hybrid nested U-shaped network and hybrid attention module enables the network to capture small,dense and adhesive nuclei,and greatly extract complex and diverse nuclear boundaries at the same time.We used the images from breast organs in MoNuSeg and 20 invasive breast cancer images from Jiangsu Province People’s Hospital to construct the invasive breast cancer segmentation dataset and trained the network to get 0.8875 F1 and 0.6419 AJI scores.After the optimization of nuclear segmentation,the prediction accuracy and sensitivity of the nuclear atypia scoring system are improved by 1.76% and 1.14%,proving it an important tool to realize the automatic grading of histology.