Book Description
To diagnose breast cancer, currently, a radiologist uses a computer-aided diagnosis system which requires them to preselect a region of interest (ROI) as an input for analysis. Breast imaging reporting and data system (BI-RADS) is a standardized reporting process to categorize breast cancer, which is based on several features of the lesion. The BI-RADS scale is based on ultrasound images, which makes the quality of the diagnosis highly dependent on the physician's experience. To minimize human error, we propose solutions based on densely connected deep convolutional neural networks. This thesis discusses various networks based on the U-Net architecture, DenseNet, attention gates, and Mask R-CNN to do semantic segmentation of the lesions from the Breast Ultrasound (BUS) images. Firstly, regular convolution blocks are replaced by dense blocks inside the U-Net (U-DenseNet), to support the learning of intricate patterns of the BUS image which is usually noisy and contaminated with speckles. This resulted in a better performance comparing to the U-Net model, with an F-score of 0.63. Then, attention gates are used in conjunction with U-DenseNet (Attention U-DenseNet) to eliminate the requirement of an explicit localization module. This resulted in a much better improvement comparing to the U-DenseNet with an F-score of 0.75. Thirdly, the previously deduced architecture, Attention U-DenseNet is used as a backbone for the Mask R-CNN architecture, which achieves an F-score of 0.76. Finally, a per-image weighted binary cross-entropy loss function is employed, as the area of the region of interest is usually small.