Mammography segmentation with maximum likelihood active contours Peyman Rahmati a,⇑ , Andy Adler a , Ghassan Hamarneh b a Dept. of Systems and Computer Engineering, Carleton University, ON, Canada b School of Computing Science, Simon Fraser University, BC, Canada article info Article history: Received 16 November 2010 Received in revised form 16 April 2012 Accepted 2 May 2012 Available online xxxx Keywords: Active contour models Computer-aided diagnosis Level sets Maximum likelihood Mammography abstract We present a computer-aided approach to segmenting suspicious lesions in digital mammograms, based on a novel maximum likelihood active contour model using level sets (MLACMLS). The algorithm esti- mates the segmentation contour that best separates the lesion from the background using the Gamma distribution to model the intensity of both regions (foreground and background). The Gamma distribu- tion parameters are estimated by the algorithm. We evaluate the performance of MLACMLS on real mam- mographic images. Our results are compared to those of two leading related methods: The adaptive level set-based segmentation method (ALSSM) and the spiculation segmentation using level sets (SSLS) approach, and show higher segmentation accuracy (MLACMLS: 86.85% vs. ALSSM: 74.32% and SSLS: 57.11%). Moreover, our results are qualitatively compared with those of the Active Contour Without Edge (ACWOE) and show a better performance. Further, the suitability of using ML as the objective function as opposed to the KL divergence and to the energy functional of the ACWOE is also demonstrated. Our algo- rithm is also shown to be robust to the selection of a required single seed point. Ó 2012 Elsevier B.V. All rights reserved. 1. Introduction Mammography has a demonstrated potential for increasing sur- vival rates through early detection of intangible tumors and small lesions (Rangayyan, 2005). Digital mammography uses X-rays to project structures in the 3D female breast onto a 2D image (Rangayyan, 2005). Breast cancer is the leading type of cancer in women and the second most fatal (American Cancer Society, 2006). Tumors ap- pear as medium-gray to white areas on digital mammograms (Egan, 1988) and their shapes are described by standardized keywords (The Mosby Medical Encyclopedia, 1992) grouped as oval, irregular, lobulated, or round, whereas their margins are expressed as circum- scribed, obscured, ill-defined or spiculated. Spiculated lesions are most often associated with cancerous pathologies (Demirkazık et al., 2003), and the presence of irregularly-shaped masses and spicules increases the probability of malignancy (Jeske et al., 2000). In addition to the variable tumor shape and appearance, the inherently noisy nature of digital mammograms, low contrast of suspicious areas, and ill-defined mass border make lesion seg- mentation an important and challenging problem. The current standard of breast cancer detection using mammo- gram is ‘‘single reading of mammograms’’ in the United States, whereas in many European countries ‘‘the double reading’’ is the standard (Gilbert et al., 2008). A current study in 2008 shows that the breast cancer detection rate using single reading with a mam- mography Computer Aided Detection (CAD) system could be sim- ilar to that of the two readers without computer assistance (Gilbert et al., 2008). In this study 31,057 women undergoing routine screening mammography at three centers in England were ran- domly assigned to double reading, single reading with CAD, or both (Gilbert et al., 2008). The cancer detection rate for the two stan- dards was measured to be equivalent. The double reading detected 199 of 227 cancers (87.7%), while the single reading with CAD de- tected 198 of 227 cancers (87.2%) (Gilbert et al., 2008). The single reading with CAD can be concluded to be an alternative to the dou- ble reading (Gilbert et al., 2008). Hence, the single reading with CAD decreases the image review process time while retaining the same accuracy as the double reading. Although mammography CAD system involves a preprocessing stage, a segmentation stage, and a classification stage, the success of clinical adoption of CAD depends mainly on the accuracy of the mammography segmentation algorithm (Baker et al., 2005). Nevertheless, there is no specific value of segmentation accuracy that guarantees the success of the overall CAD system. One study considers three different levels of segmentation accuracy for mam- mogram: near perfect, acceptable, and unacceptable (Baker et al., 2005). The unacceptable segmentation leads to a false positive decision (Baker et al., 2005). According to our results, the proposed segmentation method provides acceptable level of segmentation accuracy with an overall accuracy of 86.85%, offering a highly accu- rate segmentation stage for a CAD system. The developed segmen- tation method can be applied for treatment planning, disease 1361-8415/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.media.2012.05.005 ⇑ Corresponding author. Tel.: +1 613 866 2021; fax: +1 613 520 5727. E-mail addresses: prahmati@sce.carleton.ca, peyman.rahmati@gmail.com (P. Rahmati), adler@sce.carleton.ca (A. Adler), hamarneh@sfu.ca (G. Hamarneh). Medical Image Analysis xxx (2012) xxx–xxx Contents lists available at SciVerse ScienceDirect Medical Image Analysis journal homepage: www.elsevier.com/locate/media Please cite this article in press as: Rahmati, P., et al. Mammography segmentation with maximum likelihood active contours. Med. Image Anal. (2012), http://dx.doi.org/10.1016/j.media.2012.05.005