Whole breast ultrasound (US) is a new advanced screening technique for detecting breast abnormalities. Because several images are required to cover the entire breast region, a computer-aided diagnosis (CAD) system is employed to stitch the images together. Considerable time is needed to examine the assembled images manually, and so the CAD system is also employed as an automated diagnosis tool allowing for real time diagnosis. However, as with other automated diagnosis techniques, a reference point could be useful in defining the breast region, further reducing the tumor detection time, and also as a register for assembling the whole breast image. In automatic whole breast US, the ribs can be used as the pivotal landmark in the same way the pectoral muscle is a landmark in mammography.
Because automatic whole breast US is a very new technique, there is very little literature related to rib detection in breast US images. However, potentially suitable techniques have been developed for research on vessel detection and bone segmentation. Frangi et al1 use eigenvalue analysis of Hessian matrices to develop a vessel enhancement filter that can be applied to three dimensional visualisation modalities such as computed tomography angiography and magnetic resonance angiographies. The Hessian matrix represents the second order structures of intensity variations for each pixel of an image. The principle is to compute the curvature at each point in the image by extracting the eigenvalues (λ1, λ2, λ3) of the eigenvectors in the Hessian matrix. The ratios of the three eigenvalues obtained can be used to distinguish between tube, blob, or sheet−like structures. We adapted this technique to identify and localize the ribs in whole breast ultrascans.
As the ribs are hypoechoic, there is acoustic shadowing under them due to the inability of the sound wave to pass through. Thus, the best approach in identifying the ribs is to locate the rib shadowing in the images. Rib shadowing in the 3D whole breast US images behaves like a sheet-like structure. For a sheet-like structure, |λ3| is greater than |λ1| and |λ2|. A darkened sheet-like structure is expected to have a positive λ3 and the sum of the eigenvalues should be greater than that of the eigenvalues in background regions. Thus, a sheetness function can be derived from the three eigenvalues.
Figure 1. The screen of the SomoVu ViewStation.
The proposed method of detecting the ribs consists of five main stages. First, the breast US images are processed by subsampling to reduce the image size and length of time to analyze each image. The eigenvalues of the Hessian matrix are calculated and processed by the sheetness function to enhance the sheet-like structures and attenuate other structures. Next, the orientation thresholding method is applied to segment the structures. Known characteristics of ribs in breast US images are then applied to identify and remove non-rib components in the segmented sheet-like structures. To achieve this purpose, the connected components are extracted and labeled, and characteristics such as orientation, length, and radius are calculated. These criteria are then applied to eliminate the non-rib components.
Figure 2. (a) A slice of rib in the scanned image. The dotted line indicates the manually drawn top boundary of the rib. (b) The slice of detected rib. (c) The original whole breast image in 3D view. (d) The 3D image including the detected ribs.
Whole breast US images were acquired by the SomoVu ScanStation (U-system, San Jose, CA), an automated full-field breast US scanner. The image generated by the SomoVu ViewStation is shown in Figure 1. In our experiments, the position of a rib and the length of the rib projected to the x-z plane are used to verify the accuracy of the proposed system. The difference between the top boundary of the rib determined manually and the top boundary detected by our proposed system is used to evaluate the performance. Figure 2 shows the ribs identified in the whole breast image. A total of 65 ribs from 15 cases were analyzed. 62 ribs were detected with our method, giving a success rate of 95.38%. The accuracy of detecting the correct position, allowing a maximum difference of 5mm, is 87.10 % and the accuracy of detecting the correct length, allowing a maximum difference of 10mm, is 85.48 %.
The method can successfully detect the ribs, but estimates of the position and length of each rib need to be improved. We also found that the shadow under the nipple will affect the computed results. Hence, shadow reducing methods could be adopted in order to improve the rib detection system in the future.
This work was supported by the National Science Council, Taiwan, Republic of China, under Grant NSC95-2221-E-002 -447-MY3.
Department of Computer Science and Information Engineering/Graduate Institute of Biomedical Electronics and Bioinformatics
National Taiwan University