Research On Lesion Detection For Mammograms

Breast cancer has become a leading disease threatening women's health according to data from the World Health Organization,new cases have been increasing year by year in the world,seriously endangering women's' physical and mental health.Clinical studies have shown that early detection and treatment can effectively reduce the mortality of breast diseases.As the first choice for early screening and detection of breast diseases,mammography has good sensitivity and specificity;moreover,it is simple in operation,intuitive in imaging,safe and non-invasive,so it is recognized as the most reliable tool for early prevention and diagnosis of breast diseases.As a matter of fact,masses and micro-calcification clusters are the main signs of breast cancer in mammograms,doctors can judge the severity of breast cancer based on these lesions.However,the mass and calcification suffer the appearance of low contrast,morphological diversity and irregular shape.It is very susceptible to density tissue such as pectoral muscles and glands,therefore the radiologists face great challenges in clinical diagnosis.Accurate segmentation of breast masses and micro-calcification clusters can provide a useful reference for physicians,which can effectively improve the accuracy of clinical diagnosis.This work is mainly focused on the mammograms,aiming at exploring and analyzing the dilemmas in the breast lesions detection,improving the accuracy of early diagnosis for breast cancer.After exploring and summarizing the work of domestic and foreign scholars in lesions detection,the mammography database is collected,and then we start to explore the methods of mammographic calcification and mass segmentation by digging the intrinsic properties of the lesions.The main work of our work is summarized as follows:(a)Based on the high-frequency characteristics(noise-like characteristics)of the micro-calcification of mammograms,we propose a novel micro-calcification cluster detection method based on Contourlet transform and non-linking SPCNN model.This model can realize the separation of low-frequency background and high-frequency calcification lesions in the Contourlet domain via the multi-resolution analysis theory,which is beneficial to the subsequent fine-to-coarse micro-calcification segmentation.Based on the weak coupling characteristics between calcification points,a non-linking SPCNN model dedicated to calcification detection is designed,this model has a good biological background and is in line with human visual characteristics,sharply segmenting suspicious calcifications from the complex tissue structure in mammograms.This approach is tested on the calcification images from MIAS database and is validated on the database provided by the local hospital,the experimental results demonstrate the effectiveness and practicability of the algorithm.(b)In view of the misalignment problem of contour and inaccurate segmentation due to the weak boundary and gray unevenness in the segmentation process of mammograms,an improved CV model based on bias field theory and image local-gray statistical characteristics is proposed.In the pre-processing stage,the region growing and maximum connected region labelling is used to obtain the smooth breast region,and the nonlinear un-sharp masking is used to enhance the mass;in the lesion segmentation phase,the initial contour is automatically initialized by using spiking cortical model with biological background;then the variational level set theory is introduced to solve the initialization problem.That's,the level set evolution function deviating from the symbol distance function is compensated by adding a penalty term in the level set evolution function;for the problem of uneven gray distribution for the breast image,the bias field theory based on physical imaging principle is used to realize the grayscale correction,which can avoid the missegmentation leading by gray unevenness;finally,the local region-scalable external force based on the gray statistical characteristics is calculated as the conditionally constrain to harvest the coarse-to-fine mass boundary accurately.The algorithm is tested on DDSM and MIAS databases,and the segmentation results demonstrate the effectiveness and practicability of the algorithm.(c)By comparing and analyzing the radiologist's clinical reading process,we began to construct the artificial “eye” based on the theory and framework of the human visual attention model to achieve accurate mass detection.Specifically,a saliency detection model based on adaptive semi-local contrast features is proposed,which can analyse the distribution of the target and surrounding tissue like our human eyes by using a sliding window,reducing the background interference while preserving the region of interest.The advantages of this model are listed as: a)it is an attention-related method based on semi-local regions rather than pixel levels;b)its conditional distribution estimation is achieved by integral histogram method;c)the adaptive prior probability is used for saliency calculation instead of a fixed one.These advantages above make it have good stability and robustness for both the natural and medical images.At last,the algorithm is tested on the masses images from MIAS database.The verification experiment is conducted on the mammographic database from the local hospital,and all these experimental results show that this model has great potential in clinical applications.(d)By deeply analyzing the advantages of U-Net and residual network framework,we design and implement the Res-U-Net network by optimizing the U-Net model with residual modules and batch normalization.The model is an encoder-decoder structure,in which the spatial dimension of the pooling layer is gradually reduced in the encoder stage,the target details and spatial dimensions are gradually repaired in the decoder phase,and the skip connection strategy guarantees the better fusion of global and local information.This encoder-decoder and skip connection are very helpful to the fine mass segmentation.In our experiments,a 7-layer U-Net deep residual network was implemented using 15 convolutional layers,which can effectively solve the problem of segmentation performance degradation under the deep network architecture,so the performance of the residual U-Net network is obviously higher than the original U-Net.Finally,the feasibility and effectiveness of the model have verified on the DDSM and INbreast database,experiments results show that the proposed Res-U-Net model has a better segmentation ability than the U-Net and Seg-Net model.Besides,we collected five publicly available mammographic databases and a small database from local hospitals,these five public databases mainly include the database of MIAS,DDSM,JSMIT,INbreast,and the DWBCD features database.At the same time,we establish long-term cooperative relationships with local hospitals to expand the existing databases and provide data support for clinical diagnosis and algorithm design.