Because breast cancer lesions can have a variety of irregular shapes, it might be challenging to segment them in ultrasound images. This paper first classifies lesion masks according to their circularity to improve segmentation accuracy and find shape-specific features. Deep learning models are subsequently applied to each category to customize the segmentation method to the distinct geometric properties of the lesions. The findings reveal how lesion forms affect segmentation ability and show increased accuracy, especially for defects with irregular shapes. The accuracy of breast cancer detection in ultrasound imaging is improved by this method, which combines shape analysis and deep learning.