PARALLEL DUAL-STREAM VISION TRANSFORMER NETWORK FOR BREAST CANCER CLASSIFICATION IN MAMMOGRAPHY IMAGES

Abstract

One of the major health issues facing women globally is breast cancer, and the greatest strategy to increase patient survival rates and simplify treatment is still early identification. Because it allows doctors to detect worrisome abnormalities such masses, microcalcifications, and architectural distortions inside breast tissue, mammography is regarded as the gold standard imaging technique for breast cancer screening. However, a variety of factors, such as the subtle appearance of early-stage tumors, differences in breast density, and the existence of structural structures that may visually resemble malignant lesions, make it difficult to appropriately interpret mammograms. Sometimes, even seasoned radiologists fail to notice minor irregularities or mistake benign patterns for malignant resultsBecause of these challenges, a lot of research has been done on computer-aided diagnostic systems that can assist radiologists by providing automated analysis and decision support.In recent years, deep learning has demonstrated remarkable achievements in medical image processing, particularly for radiology classification and detection applications. Convolutional neural networks have long dominated this discipline because of their ability to extract hierarchical visual information from images. However, because CNN-based models often rely on local receptive fields, they may have trouble capturing long-range spatial correlations across huge medical images, such as mammograms.More importantly, several studies have shown that deep learning models can be impacted by shortcut learning, where the network learns to rely on non-medical cues like image borders, scanner artifacts, or high-contrast anatomical structures rather than the actual pathological features associated with disease. This habit may lead to low reliability but deceptively great accuracy during training when applied in real clinical settings. Thus, there is an increasing need for more robust architectures that can capture both global anatomical context and fine-grained disease features while avoiding bias toward irrelevant visual patterns.