Mass segmentation from digital mammograms using deep learning methods

Breast cancer is one of the main causes of cancer deaths among women. Early diagnosis is known to increase the survival rate of patients. Mammography may enable early detection of breast cancer. Digital mammography is a commonly accepted imaging modality for breast cancer screening. Abnormalities can be found from mammograms using computer-assisted methods. In recent years, deep learning research has gained momentum in computer-aided breast cancer diagnosis. As a result of these studies, when computer aided systems and radiologists work together, the system performance is increased in comparison to the radiologists' performance. Mass segmentation is an important and challenging step in the detection and diagnosis of breast cancer. Mass segmentation is one of the important steps in computer aided systems that involve identifying the location and boundaries of masses in mammography images. In this thesis, mass segmentation from whole digital mammography is investigated by using deep learning. INbreast digital mammogram data set was used for deep learning training. U-Net and its variants were used for mass segmentation performance investigation. The mass segmentation performance was tested for different models and the model with the best performance was determined. The Dice similarity coefficient that measures the intersection of the ground truch and the predicted region was used for performance evaluation. In case of training from scratch with U-net architecture, better results were obtained than other U-net variant methods. U-net performance values were compared with the ones in the literature.