Breast cancer classification and segmentation framework using multiscale CNN and U-shaped dual decoded attention network

Breast cancer is a mostly diagnosed deadly disease with a high mortality rate that can effectively be cured by early diagnosis and proper treatment. Ultrasound imaging modality is being utilized at a large scale for early diagnosis of this disease due to quick results and cheapness. Manual diagnosis of breast cancer is a laborious task that needs experts and is also subjective which necessitates the accurate diagnosis solution of this disease with help of the computer. This work proposed a combined framework for the classification and segmentation of breast cancer to automate manual diagnosis. A multiscale classification model for the classification of breast cancer into three classes is developed. For the segmentation task, an autoencoder-based U-shaped DDA-Net segmentation model consisting of a dual-decoded attention mechanism is proposed. For experiments two publically available datasets namely BUSI and UDIAT are utilized. Before the classification task feature selection method of the whale optimization algorithm is implemented and for the classification task different machine learning algorithms are utilized. The proposed method achieved the segmentation dice of 88.68% on the UDIAT dataset and dice of 87.95% on the BUSI dataset with a classification accuracy of 97.89% with a precision score of 97.9%. Results validated that the proposed combined classification and segmentation technique can reliably be implemented to accurately diagnose breast cancer at its initial level.