Breast　cancer　is　one　of　the　leading　causes　of　death　in　women　worldwide.　Therefore,　ultrasound　examination　has　become　an　important　method　of　detecting　breast　tumors.　However,　given　the　special　features　of　ultrasonic　imaging,　lesion　segmentation　is　a　challenging　task　in　computer-aided　diagnosis　systems.　In　this　study,　we　proposed　a　complex　and　automated　approach　to　segment　breast　ultrasound　images.　In　the　preliminary　contour　selection,　an　efficient　method　was　performed　by　preprocessing　of　breast　ultrasound　images,　selecting　the　iterative　threshold,　filtrating　candidate　areas,　and　ranking　remaining　areas　to　confirm　the　region　of　interest　(ROI).　After　the　selection　of　the　ROI,　a　seed　point　could　be　determined.　Then,　region　growing　started　from　the　selected　seed　to　obtain　a　preliminary　contour　that　will　serve　as　the　intermediate　result.　Finally　a　novel　and　improved　level　set　algorithm　was　proposed　to　confirm　the　final　contour,　combined　with　global　statistics,　local　statistics,　and　region-based　energy　constraint.　The　proposed　algorithm　was　tested　on　a　database　of　44　breast　ultrasound　images,　and　the　experimental　results　proved　high　accuracy.　Compared　with　the　classic　Chan-Vese　model,　the　proposed　method　increases　the　similarity　rate　and　reduces　the　error　rate.