Segmenting And Detecting Algorithm Study Of Breast Mass Based On Digital Mammograms

Mammographic imaging is especially valuable as an early detection tool for breastcancer. People have been proposed the application of computer technology to assist theclinicians with breast cancer detection and diagnosis. The thesis summarizes theauthor's research in computer—aided diagnosis on masses on mammograms, includingmammogram preprocessing, automated seeded mass segmentation and fast detection ofregion of interest.Breast boundary extracting algorithm and peripheral breast gray value correctingalgorithm were proposed in the mammogram preprocessing step. These two algorithmsaimed at extracting the breast—air interface on mammograms and enhancing the darkportions due to the reduction of thickness along the peripheral breast. The breastboundary extracting algorithm based mainly on morphological operations, especially thewatershed transformation and the boundaries from extracting algorithm were evaluatedby a clinician. Experimental results demonstrate the proposed boundary extractionalgorithm to be accurate and effective.With the application of computer image processing techniques, automated seededmass segmentation gained an optimal approximation from a seeded pixel within themass region selected manually. The author first proposed a mass model and describedthe definition of mass boundary. Then the gray—value—weighted region dilatingalgorithm and centrifugal gradient index were employed to determine the optimalboundary. We compared the segmentation of our algorithm with the mass boundarymanually drawn by clinicians. Experimental results demonstrated that the performanceof the proposed algorithm was similar to the performances of those segmentationschemes proposed by other researchers. In addition, the author also proposed the seededmatrix evaluation method to estimate the seeded pixel robustness of the segmentingalgorithm.The fast detecting algorithm of region of interest aimed at detecting possible massregions from the entire breast region on mammograms. The author employed themultiscale generic neighborhood operator to calculate the gradient of image pixels. Then,those possible mass regions were determined by analyzing the gradient orientations. Theevaluation method was a comparison between the performances of detecting algorithmand those of clinicians. Experimental results demonstrated that a sensitivity ofapproximately 90% was reached at an average level of five false positive per image.All three researching steps described above partially implement the function of acomplete computer—aided diagnosis system on breast mass.