Research On Artificial Intelligence-assisted Diagnosis Method And Clinical Application Using Mammographic Images

According to oncology-related data published by the International Agency for Research on Cancer(IARC),breast cancer has replaced lung cancer as the number one cancer in the world.Although breast cancer incidence continuously increases,thankfully,the overall mortality rate decreased by 38%from 1989 to 2014,partly attributed to breast cancer screening.The early detection,early diagnosis,and early treatment of breast cancer could improve the survival rate significantly.Imaging tests,the most critical examination of breast cancer screening,depending on doctors’experience heavily.Computer-aided detection(CAD),based on deep learning,can reduce the dependence on radiologists’ experience in the diagnosis and help improve the diagnostic rate.The CAD could also reduce the omission diagnosis rate and shorten the time of diagnosis.There are various screening methods for breast cancer screening,such as Full-field Digital Mammography(FFDM),Digital breast tomosynthesis(DBT),breast ultrasound(US),Computed Tomography(CT),Magnetic Resonance(MR),and nuclear medicine.The rise of radiomics and deep learning(DL)provided new opportunities for CAD,improving CAD performance based on different image modalities.However,there are still a series of problems in the clinical application need to be solved.Firstly,the classification of breast composition in mammography is subjective and time-consuming.The breast composition classification varies from doctor to doctor due to the difference in diagnostic experience,especially in term b,heterogeneously dense,and term c,extremely dense,in BI-RADS 5th edition.The classification in BI-RADS 5th edition depends on radiologists’ subjective decision instead of glandular percentage,resulting in low accuracy and consistency.Furthermore,the algorithm mainly focuses on the glandular percentage,which is hard to satisfy the clinical need.Moreover,it is common to obtain two projections for each breast during mammography,craniocaudal view(CC)and mediolateral oblique view(MLO),providing more information than a single projection.In clinical,radiologists match the same region of interest(ROI)on CC and MLO images by observing corresponding special location and feature similarity.However,this could be a challenging mission with gland cover and breast distortion.Therefore,it is valuable to help radiologists identify the corresponding ROI in CC and MLO images,which could help detect and diagnose breast cancer.Thirdly,calcification,cystic masses,and solid masses are essential features in breast tumor diagnosis.Calcifications are more easily detected on FFDM and DBT,while the cystic or solid masses are observed better by ultrasound.However,both FFDM and DBT bring the risk of radiation.If FFDM and DBT images could be generated from ultrasound images,more information could be provided to improve radiologist’s diagnostic accuracy.Finally,radiologists make a qualitative diagnosis by observing the density,margin,size,morphology of the mass and its accompanying signs(e.g.,skin changes,structural changes of surrounding trabeculae,lymph node morphology and density,etc.)on the mammogram.The shape is considered one of the essential features in predicting malignant breast lesions.The density and texture within the lesion and normal tissue adjacent to the tumor(NAT)also play an important role in malignant lesion prediction.However,the accuracy of a breast cancer diagnosis is affected by the experience,leading to misdiagnosis due to experience and low diagnostic level.The tumor and NAT-based joint deep learning model could fuse useful information and improve classification performance.In this study,we explore the application of machine learning and deep learning in breast density classification on a mammogram,dual-position region of interest matching,synthesized lesion of digital breast tomosynthesis from ultrasound,and fusion method-based lesion classification.1)A deep learning-based mammographic difficult-to-distinguish breast density classification.According to BI-RADS 5th edition,the most difficult breast density to distinguish,term b,heterogeneously dense,and term c,extremely dense,were selected to establish the model.A CNN-based method is used to auto classify term b and term c,and transfer learning,ensemble learning,and visualization method were applied to improve the model performance.The CNN-based ensemble model achieved a sensitivity of 0.82,a precision of 0.89,and an AUC of 0.95;the ensemble model using transfer learning had a sensitivity of 0.88,a precision of 0.93,and an AUC of 0.99.The CNN-based ensemble model achieved high performance and consistency,comparable to that of doctors with middle seniority and high seniority.2)Corresponding region of interest matching method on craniocaudal and mediolateral oblique view mammograms based on BPNN and Faster R-CNN.We proposed a BPNN and Faster R-CNN-based model to match the ROI in the dual-position mammogram.Unlike the traditional method,this study established the model based on the arc-type or straight-type matching strip by BPNN using the correlated geometry and structure of the compressed breast on the corresponding second-view mammograms could decrease the matching area and increased the detection sensitivity in CC and MLO images.The corresponding matching ROI was located by Faster R-CNN using the mass mammograms’ image content on the second-view image.Faster R-CNN+VGG16 achieved a mean distance error of 4.58 mm in 50%and 75%overlaps and accuracy of 87.04%and 85.7%,respectively.Compared with the conventional method,the proposed method achieved higher accuracy and less time consumption.3)Dual attention cycleGAN based synthesized lesion of digital breast tomosynthesis and ultrasound generation.We propose an approach to synthesizing DBT images from ultrasound images using dual attention CycleGAN.This study constructed a convolutional network-based discriminator architecture,which predicts patch labels instead of a conventional numerical label to distinguish real and fake images.Different U-net-like architertures were applied and validated for the generator network,and dual attention U-net was adopted for experimental performance.The DA-Unet was suitable for CycleGAN and achieved a low freshet inception distance(FID)of 208.7.According to the radiologist,the DA-Unet achieved a high score close to real DBT,with an accuracy of 0.83,a sensitivity of 0.71,and a specificity of 0.90.Based on objective and subjective evaluation,the image quality of synthesized DBT is comparable to real DBT.4)A joint deep learning model for predicting breast lesion classification.This study fused the deep learning model and machine learning model and intergrate information of lesion shell and kernel to simulate a radiologist’s decision-making process using DBT lesions.The classification performance was improved after information fusion,achieving an AUC of 0.91 after Shell&Kernel fusion.Compared with radiologists,the fusion model’s performance was better than that of junior radiologists and comparable to middle seniority radiologists,and lower than senior radiologists.With the assistance of the fusion model,most radiologists could improve their diagnostic performance in breast masses.In this study,we studied several artificial intelligence methods such as UNet,VGG-16,VGG-19,ResNet-50,DenseNet,and support vector machine,and applied them to difficult-to-distinguish density classification,the corresponding region of interest matching in two views,ultrasound-mammography synthesized lesion generation and breast lesion classification using joint deep learning model.The excellent performance of AI in the above application proves the powerful ability and unlimited potential of AI,which will continue to play an essential role in applying breast image assist diagnosis.