A robust method for segmenting pectoral muscle in mediolateral oblique (MLO) mammograms

• Published in: International journal for computer assisted radiology and surgery Vol. 14; no. 2; pp. 237 - 248
• Main Authors: Yin, Kaiming, Yan, Shiju, Song, Chengli, Zheng, Bin
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.02.2019; Springer Nature B.V
• Subjects: Algorithms; Automation; Breast Density; Breast Neoplasms - diagnostic imaging; Computer Imaging; Computer Science; Contours; Data analysis; Female; Health Informatics; Hough transformation; Humans; Imaging; Iterative methods; Least squares method; Mammography - methods; Medicine; Medicine & Public Health; Muscles; Original Article; Pattern Recognition and Graphics; Pectoralis Muscles - diagnostic imaging; Radiographic Image Interpretation, Computer-Assisted - methods; Radiology; Segmentation; Shape; Surgery; Vision; Pectoral muscle; Automated segmentation; Computer-aided diagnosis; Mediolateral oblique mammograms
• ISSN: 1861-6410; 1861-6429; 1861-6429
• DOI: 10.1007/s11548-018-1867-7

Purpose Accurately detecting and removing pectoral muscle areas depicting on mediolateral oblique (MLO) view mammograms are an important step to develop a computer-aided detection scheme to assess global mammographic density or tissue patterns. This study aims to develop and test a new fully automated, accurate and robust method for segmenting pectoral muscle in MLO mammograms. Methods The new method includes the following steps. First, a small rectangular region in the top-left corner of the MLO mammogram which may contain pectoral muscle is captured and enhanced by the fractional differential method. Next, an improved iterative threshold method is applied to segment a rough binary boundary of the pectoral muscle in the small region. Then, a rough contour is fitted with the least squares method on the basis of points of the rough boundary. Last, the fitting contour is subjected to local active contour evolution to obtain the final pectoral muscle segmentation line. The method has been tested on 720 MLO mammograms. Results The segmentation results generated using the new scheme were evaluated by two expert mammographic radiologists using a 5-scale rating system. More than 65% were rated above scale 3. When assessing the segmentation results generated using Hough transform, morphologic thresholding methods and Unet-based model, less than 20%, 35% and 47% of segmentation results were rated above scale 3 by two radiologists, respectively. Quantitative data analysis results show that the Dice coefficient of 0.986 ± 0.005 is obtained. In addition, the mean rate of errors and Hausdorff distance between the contours detected by automated and manual segmentation are FP = 1.71 ± 3.82%, FN = 5.20 ± 3.94% and 2.75 ± 1.39 mm separately. Conclusion The proposed method can be used to segment the pectoral muscle in MLO mammograms with higher accuracy and robustness.