Computer Aided Screening Of Breast Cancer In The Whole Breast Ultrasound Image

Breast cancer is the most common cancer and the leading cause of cancer-related deaths for women all over the world.Early screening and treatment of breast cancer have been shown to be useful in reducing mortality rates.Automated breast ultrasound(ABUS)is a new and promising screening modality for whole breast examination.Compared to conventional handheld ultrasound,ABUS achieves operator-independent image acquisition and can provide 3D views of the whole breast.However,reviewing ABUS images is extremely time-consuming and oversight errors could happen during the examination.Therefore,automated cancer detection in ABUS is highly expected to assist clinicians in facilitating the identification of breast cancer.The present study is aimed to detect breast cancer in ABUS and further develop a robust and accurate computer-aided screening tool.In the 3D-DUnet method,we regard the cancer screening as segmentation task in computer vision(CV).Since the basic network 3D-Unet can’t meet the sensitivity and specificity of cancer screening.So,we present the following threefold contribution.Firstly,we proposed a threshold mask optimized by threshold loss function to provide voxel-level adaptive threshold for discriminating cancer and non-cancer,thus achieving high sensitivity with low false positive(FPs).Secondly,we propose a densely deep supervision(DDS)mechanism to improve the sensitivity significantly by utilizing multi-scale discriminative features of all layers.Both dice loss and overlap loss are employed to enhance DDS performance.Thirdly,we delete a pooling layer,which can deal with the problem of missing small cancer caused by consecutive down-sample operation.Besides,equip the network with 3D dilated convolution simultaneously to expand receptive field,which can extract more features from shadow and lesions in ABUS for high sensitivity with low FPs.In the 2D-CALN method,the cancer screening is considered as detection task in CV field.The input is two-dimensional images of the transverse ABUS lesion region without down-sample.The candidate area location network(CALN)is employed to extract cancer region proposals on feature maps of the input images,and then use the regional classification and regression,identify the candidate box containing cancers and the cancers’location information.In addition,we adopt recurrent neural networks(RNN)to refine the feature maps between adjacent images,further enhance cancer screening performance.The efficacy of the proposed methods is verified on a dataset of 219 malignant tumor patients with 615 volumes and 747 cancer regions,and 144 normal healthy women with 900 volumes.Proved by the 4-fold cross-validation experiments,3D-DUnet system obtains a sensitivity of 95.12% with 0.84 FPs per ABUS volume.We can conclude that our proposed novel network with DDS and TM can provide an accurate and automatic cancer detection tool for breast cancer screening by maintaining high sensitivity with low FPs.Despite this,it has a problem of missing small cancer less than 1cm~3.Besides,2D-CALN system achieves a sensitivity of 96.13%with 2.02 FPs per ABUS volume,which shows that the model based on the higher resolution has good sensitivity,especially for small cancer.However,due to the lack of spatial information constraints,FPs are relatively higher.Subsequent research will employ multi-scale technology to combine the above two methods.In summary,a great deal of experiments have proved that provides a fully automated,accurate and efficient tool for clinical breast cancer screening,which has important clinical application and promotion value.