| At present,breast cancer is the first fatal cancer in women,but timely detection and treatment can effectively reduce the lethality rate of breast cancer.Mammography has become the first choice for the diagnosis of breast diseases because of its simplicity,reliability and noninvasive characteristics.In the clinical process,radiologists will give corresponding diagnostic opinions according to the patient’s breast mammograms,but the manual mammogram reading process is not so easy.As a matter of fact,even experienced radiologists may make mistakes,resulting in missed diagnosis and misdiagnosis.Therefore,computer-aided diagnosis system came into being.The main purpose of CAD system is to provide clinicians with diagnostic opinions for reference,simplify the clinical mammogram reading process and reduce the phenomenon of missed diagnosis and misdiagnosis.The whole diagnosis process of CAD system based on mammograms is composed of several parts.In this thesis,two parts,breast density classification and full field breast mass segmentation,are deeply studied,which are summarized as follows:1.Breast density classification based on bilateral adaptive spatial and channel attention networkBreast density is an important factor to prevent breast cancer.The existing breast density classification methods can not realize end-to-end classification and can not achieve satisfactory classification accuracy.In this thesis,we propose a bilateral adaptive spatial and channel attention network（BASCNet）,which integrates the information of patients’ left and right breast and adaptively focuses on the discriminative features in spatial and channel dimensions.The proposed BASCNet has been fully verified on the two public datasets of DDSM and INBreast.The classification accuracy of five-fold cross validation is 85.10% and90.51% respectively.Our method is fully automatic and achieves better classification performance than the existing breast density classification methods.At the same time,we conducted a large number of ablation experiments to prove the effectiveness of the network topology.In addition,we also compared the effects of different views（CC and MLO）on breast density classification,and verified the effectiveness of contralateral breast information integration.In general,the proposed BASCNet has the potential for clinical diagnosis.2.Full-field breast mass segmentation based on adaptive channel and multiscale spatial context networkBreast mass is the most obvious means of cancer identification.Therefore,accurate segmentation of mass is very important.Compared with the mass-centered patch segmentationon,it is always a challenging subject to accurately segment the breast mass in the full-field mammograms,because its signal-to-noise ratio is very low,the shape,size and location of the mass are uncertain,and the network is easy to predict false positive.In this study,we propose an adaptive channel and multi-scale spatial context network（ACMSCNet）for breast mass segmentation in full-field mammograms.ACMSCNet adopts standard encoder-decoder architecture and embeds adaptive channel and multi-scale spatial context module（ACMSC module）in a multi-level way to achieve accurate mass segmentation.The proposed ACMSC module uses the self-attention mechanism to adaptively capture the discriminative context information between the channel and the spatial dimension.The multi-level embedding of ACMSC module enables the network to learn different features on multi-scale feature map.The proposed model is evaluated on two public datasets CBIS-DDSM and INBreast.The experimental results show that by adaptively capturing the context of channel and spatial dimensions,our model can effectively remove false positives,predict difficult samples,and obtain better results than the previous methods: the Dice coefficients on CBIS-DDSM and INBreast datasets are 82.81% and 84.11% respectively.We hope that our work will contribute to the CAD system for breast cancer diagnosis and ultimately improve the level of clinical diagnosis. |
PDF Full Text Request