Improved deep joint segmentation and deep hybrid architecture for breast cancer diagnosis with improved pattern extractors

Abstract

Cancer remains one of the primary contributors to illness and mortality among women globally. One in five new cases of cancer in the world occurs in the breast. The use of mammography as a radiological diagnostic tool for breast abnormalities resulted in a notable reduction in breast cancer (BC) death rates. It can be particularly challenging to identify and categorize breast tumors accurately on mammograms for a variety of reasons, such as low contrast and typical fluctuations in tissue densities. Different CAD systems have been created to help radiologists correctly classify abnormalities of the breast. Still, they lack accurate detection. This work develops a new BCD model using Deep Hybrid Architecture (DHA). Initially, the histogram equalization (HE) technique is deployed for preprocessing the input mammogram image. Then, Improved Deep Joint (IDJ) is used for segmenting the image. Further, Improved Local Gabor Binary Pattern Histogram Sequence (ILGBPHS), Median Binary Pattern (MBP) and Local Gradient Pattern (LGP) features are extracted. Finally, DHA, including Improved Deep Convolutional Neural Network (IDCNN) and LinkNet, are utilized to detect the BC. The results from DHA are averaged, and the BCD results are achieved. The proposed DHA framework attained an accuracy of 95.00% on the mini-MIAS dataset and 95.03% on the CBIS-DDSM dataset, demonstrating superior performance compared with conventional DJ, conventional LGBPHS, and DCNN+LinkNet methods.