Breast cancer is a common disease in women all around the world, however the exponential growth of CAD systems and artificial intelligence has allowed for the early detection of asymptomatic tumours. Segmentation provide information pertaining to the anatomical structure of the tumour and their tissue volume which could be of a great help to radiologists. Traditional segmentation techniques are time consuming, hence, an automated segmentation utilising deep learning algorithms could help radiologists for an early diagnosis. In this study, an attention based Unet model is employed to segment masses in mammograms. The developed model is validated on the benchmark dataset CBIS-DDSM. In this study, the impact of several loss functions are experimented with our proposed model. In comparison to the other models, the attention based Unet produced a dice score of 90.4% and a Intersection over union (IoU) of 84.9%.