Breast Cancer Classification Using a Novel Image Processing Pipeline and a Two-Stage Deep Learning Segmentation and Classification Approach

Mammography is used as a primary method for the X-ray imaging of breast regions. These mammogram images are utilized by radiologists for localization and prognosis of the mass regions present in the breast region, as either malignant or benign. We present a similar approach in our work, where a two-stage system is proposed for the localization and classification of breast mass regions using the mammogram images. First, these mammograms are passed through an image processing pipeline for the initial processing. Secondly, these processed images are fed into the proposed two-stage system for segmentation and classification. For the segmentation stage, we use the UNET segmentation architecture with EfficientNetB0, ResNet50, and MobileNet encoders without any pre-trained weights. For the classification stage, we use the VGG16 and ResNet50 pre-trained models for our task where we feed in the segmented region of interest of tumors as input and the output of the model being the pathology of the tumor. The results obtained show good accuracy in determining the pathology of the mass region in the mammogram images, with results obtained at low latency with good precision, recall, specificity, and sensitivity rates.