DEES-breast: deep end-to-end system for an early breast cancer classification

Breast cancer mortality reduction progress has halted in recent years. The mortality rate was rising, and breast cancer was the leading cause of death among women. Early diagnosis is critical in treatment since it can prevent complications and heavy pathologic therapy. Many Computer-Aided Diagnosis (CAD) systems were developed for this purpose. However, to produce more accurate findings, it must continue to be enhanced by adopting new methodologies. To efficiently handle semantic segmentation in a predicted image, we propose a novel Fully Convolutional Network (FCN) called DEES-Breast that presents an End-to-End system for an early breast cancer detection from mammographic scans. The DEES-Breast uses an encoder-decoder architecture to identify relevant features from scans at several scales and upsample them to generate the best segmentation results. The main advantage of the proposed architecture is the skip connection mode within the decoder and encoder layers, which merges high-level features encoded with low-level features decoded from the decoder. The CNN used at the encoder tries to admit relevant studies having similar contrast values using thirteen convolutional layers and three fully connected layers. Various complex preprocessing methods were carefully used to enhance the model's performance. These methods included various procedures, such as image cropping, CLAHE enhancement, artifact removal, etc., and allowed us to create a well-prepared dataset for training and testing. Geometric data augmentations were carefully integrated into the pipeline to improve generalization capabilities and reduce overfitting. CBIS-DDSM images and a private database were used to test our suggested architecture comprehensively. Quantitative criteria for evaluating segmentation outcomes, such as Dice coefficient, precision, and recall, are all above 90%, demonstrating that the proposed architecture system can differentiate functional tissues in breast mammogram images. As a result, our proposed architecture has the potential to offer the classification required to aid in the clinical detection of breast cancer while also improving imaging in other modalities of medical mammography.