Fusion of Local and Global Representations Using a Hybrid CNN–Attention Network for Breast Tumor Classification

Abstract

Diagnosis of breast cancer with the help of medical imaging is one of the critical factors of decreasing mortality, however, it is
challenging to obtain the necessary consistency and reliability in the classification due to low contrast and tissue heterogeneity and fine
distinctions between the benign and malignant patterns. Traditional computer-aided systems are mainly based on hand-crafted descriptors and
superficial classifiers, that find it hard to represent fine spatial dependencies, and that tend to be very weak in the face of a wide range of
clinical data. It is on these weaknesses that the necessity arises of automated models that can learn discriminative representations by viewing
images. The paper presents a new hybrid deep learning framework of breast cancer classification that combines the convolutional representation
learning with complementary global context modelling. First, the input images are processed through various activities such as noise reduction,
intensity normalization and adaptive contrast enhancement to reduce artificialities of the acquisition process. An optional lightweight
segmentation strategy is used to focus the lesion part and minimize the background redundancy. A convolutional backbone is used to extract
deep spatial features, and a parallel transformer-inspired attention branch is used to learn long-range structural relationships. This fused feature
space takes a local morphology and global dependencies and the aggregated embeddings are further fed into a multilayer classification module
to make benign or malignant predictions. To provide clear benchmarking, the proposed approach is justified using the Wisconsin Diagnostic
Breast Cancer dataset. Through experimental assessment, the hybrid design has been found to devise greater discriminative capacity and a high
overall accuracy rate of 97.2 and thus competitive to most of the recent practices. The framework emphasizes the usefulness of integrating
local and global learning paradigm to provide reliable and scalable breast cancer decision support.