Attention : Cascaded UNets with Modified Skip Connection for Breast Tumor Segmentation

Amidst of all the tumors, breast tumor is the main source for a high incidence mortality rate among women. In the past decades, this mortality rate has tremendously reduced by early detection and diagnosis of the tumors using different modalities. Mammograms have significantly contributed towards the detection of masses and calcifications which assists radiologists in decision making. With the advent of Artificial Intelligence (AI) techniques, several breast mass segmentation methods are devised to assists radiologists in detection of tumor masses and lesions. UNet and its variants are widely used deep learning techniques for mass segmentation in mammograms. Although, UNet exhibits an outstanding performance, it has certain limitations, where the extraction of the fine-grained features is still resilient. In this paper, a AU\(^{2}\)Net is proposed where two UNets are cascaded and are connected through skip connections. The attention mechanism and Atrous Spatial Pyramid Pooling (ASPP) are introduced to learn more contextual based features for an enhanced tumor segmentation. The proposed model is validated on Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM). The images are preprocessed and are augmented using conventional and deep convolutional Generative Adversarial Network based approaches. The proposed method achieved a 92.3% dice score and 89.6% Intersection over Union (IoU) on CBIS-DDSM dataset.