Luan Jiang PhD, Abstract Co-Author: Nothing to Disclose

Yanyun Lian, Abstract Co-Author: Nothing to Disclose

Yajia Gu MD, Abstract Co-Author: Nothing to Disclose

Xiaoxin Hu, Abstract Co-Author: Nothing to Disclose

Qiang Li PhD, Presenter: Patent agreement, General Electric Company Patent agreement, Hologic, Inc Patent agreement, Riverain Technologies, LLC Patent agreement, MEDIAN Technologies Patent agreement, Mitsubishi Corporation

We are developing a computer-aided diagnostic (CAD) scheme for breast cancer diagnosis, risk assessment, and density analysis in magnetic resonance (MR) image. This study is to develop a key technique in the CAD for accurate segmentation of the whole breast.

We obtained 37 clinical transverse breast MR scans from Shanghai Cancer Hospital in China. All images were fat-suppressed T1-weighted MR acquired with an Aurora 1.5-T MR scanner. The image size was 512×512×256 and the resolution in x, y, z axis was 0.7031 mm after linear interpolation in z axis. Sixteen breasts in 8 cases were manually delineated by a physicist and confirmed by a radiologist for objective evaluation of three-dimensional (3-D) segmentation method. We first developed techniques to determine a bounding box for a breast and to extract the major part of the breast by use of a 3-D spiral scanning method and dynamic programming. A small portion of the breast around the pectoral muscle was not included in the initial segmentation result. We then further developed a technique to delineate the chest wall and to add the missing small portion of the breast back to the major portion. A sectional dynamic programming method was designed in each 2-D slice of a 3-D MR scan to trace the upper and/or lower boundaries of the chest wall. Our method also took advantages of the continuity of chest wall across adjacent slices. The performance level of the whole breast segmentation method was subjectively observed by human readers, and was evaluated by objective indices of Dice overlap measure and volume agreement.

By subjective observation of the 37 cases, we found that the proposed method obtained good segmentation of the whole breasts. By comparing with the manually delineated region of 16 breasts in 8 cases, an overlap index of 90.5% ± 2.6% (mean ± SD), and a volume agreement of 96.2% ± 3.5% were achieved, respectively.

The fully automated method for accurate segmentation of the whole breast would be useful for developing CAD systems for risk assessment and early diagnosis of breast cancer in MRI.

Our CAD scheme could help the radiologists to quantitatively analyze breast cancer risk and to improve the accuracy and efficiency of breast cancer diagnosis.

Jiang, L, Lian, Y, Gu, Y, Hu, X, Li, Q, Fully Automated Segmentation of Whole Breast in MR images Using Dynamic Programming.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13025798.html