Deep Learning Techniques for Segmenting Breast Lesion Regions and Classifying Mammography Images

Nam V. Nguyen, Hieu Trung Huynh, Phuc-Lu Le

Abstract: Breast cancer is currently one of the leading causes of death in many countries worldwide. Detecting breast masses early can provide higher chances of survival for patients. However, determining and segmenting benign or malignant breast masses is becoming a challenging issue. Currently, there are a wide range of Convolutional Neural Networks used to address breast mass segmentation and breast cancer classification issues, such as U-Net, SegNet, Mask R-CNN, for segmentation, and Convnet, CNN, ResNet, for classification. However, these solutions are still not effective enough. Therefore, we have solved this problem by applying modern model called Segment Anything Model to predict breast tumor segmentation masks to help doctors identify and evaluate breast tumors and two models EfficientNet B0 combined with Focal Loss and Vision Transformer base to classify breast images as benign or malignant. The experimental results show those modern models achieved high performance with an Intersection over Union score of 96.59% on the CIBS-DDSM dataset. Additionally, the classification model achieved an accuracy of 100% and F1-scores of 100% on the DDSM dataset, outperforming other models. Our technique helps support doctors in identifying breast masses in images and provides reliable predictions for diagnostic purposes, thus improving the effectiveness of breast cancer detection.