Multi-channel Deep Neural Network Synergic Fusion For Breast Mass Classification Using Digital Breast Tomosynthesis

Breast cancer has become one of the most common cancers that threaten women’s physical and mental health in today’s society.Early detection and treatment is the key to improving breast cancer survival and restoring patients’ health.Digital breast tomography(DBT)as a new breast examination method can effectively improve the screening rate of breast cancer.In recent years,the effective combination of medical images and artificial intelligence has promoted the development and application of computer-aided diagnosis(CAD)systems.Deep convolutional neural networks(DCNN)can learn existing image data to identify abnormalities in undiagnosed images.This auxiliary tool helps Doctors have improved the speed and quality of diagnosis.Although a large number of studies have achieved certain results in other breast diagnosis imaging,there is still room for improvement in the classification of breast masses in DBT.Therefore,for the classification of benign and malignant breast cancer masses on DBT images by DCNN,this paper proposes a multi-path synergic fusion(MSF)deep neural network model.To verify the performance of the model,we retrospectively collected 441 patients who had undergone DBT,extracted the regions of interest(ROIs)covering the benign/malignant breast mass,and then extracted three multifaceted representations of the breast mass(gross mass,overview,and mass background)from the ROIs.The main reserch contents of this paper are as follows:1)A multi-scale multi-level features enforced DenseNet(MMFED)is proposed to process the three images independently.The network structure is based on DenseNet and branches are added to extract multi-scale multi-level features.According to the final classification results on the three images,the advantages of MMFED over DenseNet are comprehensively explained.2)Four decision fusion strategies including the plurality voting(PV),weighted fusion(WF),stacking,and decision template(DT)methods are used to fuse the output of the three sub-models at the decision level to generate the final prediction.We also compared various fusion frameworks to verify the advantages of MSF in the classification of benign and malignant breast masses.Experimental results show,the MMFED was observed to be superior to the original DenseNet,and multiple channel fusions in the MSF outperformed the singlechannel MMFED and double-channel fusion with the best classification scores of area under the receiver operating characteristic(ROC)curve(87.03%),Accuracy(81.29%),Sensitivity(74.57%),and Specificity(84.53%)via the weighted fusion method embedded in MSF.The decision level fusion-based MSF was excelled(in terms of the receiver operating characteristic(ROC)curve)than the feature fusion model(p<0.05),the channel fusion model(p<0.04),and the end-to-end model(p<0.004).According to the above results,integrating multifaceted representations of the breast mass tends to increase benign/malignant mass classification performance and the proposed methodology was verified to be a promising tool to assist in clinical breast cancer screening.