The Research On Automatic Detection Of Clustered Microcalcifications In Digital Breast Tomosynthesis

Breast cancer is one of the malignant cancers which threatens to women’s health.And there is an increasing incidence among most of the countries in the world.The survival rate of breast cancer patients is related to the detection period of the lesion.The earlier the breast cancer is diagnosed,the better the prognosis will be.Microcalcification cluster is an important index of early malignant breast tumors in clinical diagnosis.Therefore,accurate detection of calcification cluster is very important for the diagnosis and reasonable treatment planning of breast cancers,which contributes to reduce the mortality of breast cancer patients.Nowadays classical breast x-ray photography is a common way to detect breast cancer because of its effectiveness and simplicity.But the overlapping of breast tissues may leads to undetected lesion and microcalcification cluster.The emergence of new imaging technology,digital breast tomosynthesis(DBT),effectively overcomes the problems.especially it significantly improves the detection rate of calcified clusters.However,DBT images often have a large number of sectional image data which greatly increases the workload of radiologists.So it is urgent to study the automatic detection algorithm of calcification clusters in DBT images to assist doctors in the diagnosis and treatment of breast cancer.In recent years,most researchers only focused on DBT tomographic images or projected views,ignoring the three-dimensional structure information within DBT images.Based on the three-dimensional characteristics of DBT images,this study proposes a three-dimensional Hessian matrix-based calcification enhancement algorithm to pre-screen calcification points.After that,combined with clinical prior knowledge,microcalcification clusters are segmented by three-dimensional spatial clustering of candidate calcification points,and then the texture features of microcalcification clusters are extracted to train random forest classifiers to reduce the false positivity of microcalcification clusters.Specific work includes:(1)Pre-processing of images:self-adaptive window is applied to enhance the image quality which makes the contrast of microcalcification points effectively increased.And the breast muscle on MLO(mediolateral oblique)is segmented for decreasing the false positivity and subsequent computation assumption based on Hough transform.This step lays the necessary bases for the detection of calcification points.(2)Pre-screening of calcification points:A Laplacian pyramid-based multi-scale bilateral filtering denoising algorithm is proposed,which can reduce noise interference while maintains the edge features of microcalcification points and effectively improve the pre-screening quality of microcalcification points.Considering that the calcification points in DBT images have the characteristics of three-dimensional spatial morphology and high brightness and contrast,two parallel filters are used to enhance the microcalcification points of the pre-processed images.One is the spherical structure discrimination enhancement based on multi-scale Hessian matrix,and the other is the enhancement using combined linear filters to suppress organizational structure background and boost the signal-to-noise ratio(SNR)of calcification point information on slice images.After that the two filters are fused to remove the low-frequency background to make the potential information of calcification point more prominent.Finally,threshold-based 26-neighborhood region growing algorithm is utilized to locate the potential microcalcification point in the weighted body enhanced by filters.And the final candidate microcalcification points are pre-screened using region growing algorithm based on the potential microcalcification point center from signal-to-noise ratio enhancement results.which obtains the information of microcalcification point necessary for subsequent extraction of microcalcification clusters area.(3)Detection of calcification clusters:Based on the results of last step,this step uses clustering method to segment calcification points area.Then,the random forest classifier is trained to reduce false positives according to the texture and statistical features of the maximum density projection image in three-dimensional region of each calcification cluster.Finally,the performance of the model is evaluated by FROC curve and non-parametric method.For case-based detection,a sensitivity of 85%was achieved at an FP rate of 0.76 per DBT volume.This study shows that using three-dimensional information of DBT images to detect microcalcification clusters can get accurate detection results,and through simple and effective classifier algorithm,the false positive of calcified clusters can be greatly reduced,which fully confirms the feasibility of the algorithm,and provides a new method for assisting doctors in early detection of breast cancer.