Images data practices for Semantic Segmentation of Breast Cancer using Deep Neural Network

Image data in healthcare is playing a vital role. Medical data records are increasing rapidly, which is beneficial and detrimental at the same time. Large Image dataset are difficult to handle, extracting information, and machine learning. The mammograms data used in this research are low range x-ray images of the breast region, which contains abnormalities. Breast cancer is the most frequently diagnosed cancer and ranked 9th worldwide in breast cancer-related deaths. In Pakistan 1 in 9 women expected to have breast cancer at some stage in life. Screening mammography is the most effective means for its early detection. This high rate of oversampling is responsible for billions in excess health care cost and unnecessary patient anxiety. This research mainly focuses on the development of deep learning based computer-aided system to detect, classify and segment the cancerous region in mammograms. Moreover, the preprocessing mechanism is proposed that remove noise, artifacts and muscle region that can cause a high false positive rate. In order to increase the efficiency of the system and counter the large resource requirement, the pre-processed image is converted to 512 x 512 patches. The two publicly available breast cancer dataset are employed i.e. Mammographic Image Analysis Society (MIAS) digital mammogram dataset and Curated Breast Imaging Subset of (Digital Database for Screening Mammography) (CBIS-DDSM). The two states of art deep learning-based instance segmentation frameworks are used, i.e. DeepLab and Mask RCNN. The pre-processing algorithm helps to increase the area under the receiver operating curve for each transfer learning method. The fine tuning is performed for better performance, the area under the curve was equal to 0.98 and 0.95 for mask RCNN and deep lab respectively on a test set of 150 cases. However, mean average precision for the segmentation task is 0.80 and 0.75. The radiologists accuracy ranged from 0.80 to 0.88. The proposed research has the potential to help radiologists with breast mass classification as well as segmentation of the cancerous region.